Contrastive learning has become a fundamental approach in both uni-modal and multi-modal frameworks. This learning paradigm pulls positive pairs of samples closer while pushing negatives apart. In the uni-modal setting (e.g., image-based learning), previous research has shown that the strength of these forces can be controlled through the temperature parameter. In this work, we propose Multi-Modal Temperature and Margin Schedules (MM-TS), extending the concept of uni-modal temperature scheduling to multi-modal contrastive learning. Our method dynamically adjusts the temperature in the contrastive loss during training, modulating the attraction and repulsion forces in the multi-modal setting. Additionally, recognizing that standard multi-modal datasets often follow imbalanced, long-tail distributions, we adapt the temperature based on the local distribution of each training sample. Specifically, samples from dense clusters are assigned a higher temperature to better preserve their semantic structure. Furthermore, we demonstrate that temperature scheduling can be effectively integrated within a max-margin framework, thereby unifying the two predominant approaches in multi-modal contrastive learning: InfoNCE loss and max-margin objective. We evaluate our approach on four widely used image-and video-language datasets, Flickr30K, MSCOCO, EPIC-KITCHENS-100, and YouCook2, and show that our dynamic temperature and margin schedules improve performance and lead to new state-of-the-art results in the field.1

Abstract Reasoning over sports videos for question answering is an important task with numerous applications, such as player training and information retrieval. However, this task has not been explored due to the lack of relevant datasets and the challenging nature it presents. Most datasets for video question answering (VideoQA) focus mainly on general and coarse-grained understanding of daily-life videos, which is not applicable to sports scenarios requiring professional action understanding and fine-grained motion analysis. In this paper, we introduce the first dataset, named Sports-QA, specifically designed for the sports VideoQA task. The Sports-QA dataset includes various types of questions, such as descriptions, chronologies, causalities, and counterfactual conditions, covering multiple sports. Furthermore, to address the characteristics of the sports VideoQA task, we propose a new Auto-Focus Transformer (AFT) capable of automatically focusing on particular scales of temporal information for question answering. We conduct extensive experiments on Sports-QA, including baseline studies and the evaluation of different methods. The results demonstrate that our AFT achieves state-of-the-art performance.

We introduce MEt3R, a metric for multi-view consistency in generated images. Large-scale generative models for multi-view image generation are rapidly advancing the field of 3D inference from sparse observations. However, due to the nature of generative modeling, traditional reconstruction metrics are not suitable to measure the quality of generated outputs and metrics that are independent of the sampling procedure are desperately needed. In this work, we specifically address the aspect of consistency between generated multi-view images, which can be evaluated independently of the specific scene. Our approach uses DUSt3R to obtain dense 3D reconstructions from image pairs in a feed-forward manner, which are used to warp image contents from one view into the other. Then, feature maps of these images are compared to obtain a similarity score that is invariant to view-dependent effects. Using MEt3R, we evaluate the consistency of a large set of previous methods for novel view and video generation, including our open, multi-view latent diffusion model. Code is available online: geometric-rl.mpi-inf.mpg.de/met3r/.

We introduce PersonaHOI, a training-and tuning-free framework that fuses a general StableDiffusion model with a personalized face diffusion (PFD) model to generate identity-consistent human-object interaction (HOI) images. While existing PFD models have advanced significantly, they often overemphasize facial features at the expense of full-body coherence, PersonaHOI introduces an additional StableDiffusion (SD) branch guided by HOI-oriented text inputs. By incorporating cross-attention constraints in the PFD branch and spatial merging at both latent and residual levels, PersonaHOI preserves personalized facial details while ensuring interactive non-facial regions. Experiments, validated by a novel interaction alignment metric, demonstrate the superior realism and scalability of PersonaHOI, establishing a new standard for practical personalized face with HOI generation. Code is available at https://github.com/JoyHuYY1412/PersonaHOI.

We propose a new "Unbiased through Textual Description (UTD)" video benchmark based on unbiased subsets of existing video classification and retrieval datasets to enable a more robust assessment of video understanding capabilities. Namely, we tackle the problem that current video benchmarks may suffer from different representation biases, e.g., object bias or single-frame bias, where mere recognition of objects or utilization of only a single frame is sufficient for correct prediction. We leverage VLMs and LLMs to analyze and debias benchmarks from such representation biases. Specifically, we generate frame-wise textual descriptions of videos, filter them for specific information (e.g. only objects) and leverage them to examine representation biases across three dimensions: 1) concept bias — determining if a specific concept (e.g., objects) alone suffice for prediction; 2) temporal bias — assessing if temporal information contributes to prediction; and 3) common sense vs. dataset bias — evaluating whether zero-shot reasoning or dataset correlations contribute to prediction. We conduct a systematic analysis of 12 popular video classification and retrieval datasets and create new object-debiased test splits for these datasets. Moreover, we benchmark 30 state-of-the-art video models on original and debiased splits and analyze biases in the models. To facilitate the future development of more robust video understanding benchmarks and models, we release: "UTD-descriptions", a dataset with our rich structured descriptions for each dataset, and "UTD-splits", a dataset of object-debiased test splits.1

Video Large Language Models (Vid-LLMs) have made remarkable advancements in comprehending video content for QA dialogue. However, they struggle to extend this visual understanding to tasks requiring precise temporal localization, known as Video Temporal Grounding (VTG). To address this, we introduce Number-Prompt (NumPro), a novel method that empowers Vid-LLMs to bridge visual comprehension with temporal grounding by adding unique numerical identifiers to each video frame. Treating a video as a sequence of numbered frame images, NumPro transforms VTG into an intuitive process: flipping through manga panels in sequence. This allows Vid-LLMs to “read” event timelines, accurately linking visual content with cor responding temporal information. Our experiments demonstrate that NumPro significantly boosts VTG performance of top-tier Vid-LLMs without additional computational cost. Furthermore, fine-tuning on a NumPro-enhanced dataset defines a new state-of-the-art for VTG, surpassing previous top-performing methods by up to 6.9% in mIoU for moment retrieval and 8.5% in mAP for highlight detection. The code is available at https://github.com/yongliang-wu/NumPro.

Visual in-context learning (VICL), as a new paradigm in computer vision, allows the model to rapidly adapt to various tasks with only a handful of prompts and examples. While effective, the existing VICL paradigm exhibits poor generalizability under distribution shifts. In this work, we propose test-time Visual In-Context Tuning (VICT), a method that can adapt VICL models on the fly with a single test sample. Specifically, we flip the role between the task prompts and the test sample and use a cycle consistency loss to reconstruct the original task prompt output. Our key insight is that a model should be aware of a new test distribution if it can successfully recover the original task prompts. Extensive experiments on six representative vision tasks ranging from high-level visual understanding to low-level image processing, with 15 common corruptions, demonstrate that our VICT can improve the generalizability of VICL to unseen new domains. In addition, we show the potential of applying VICT for unseen tasks at test time. Code: https://github.com/Jiahao000/VICT.

Neural networks are widely adopted to solve complex and challenging tasks. Especially in high-stakes decision-making, understanding their reasoning process is crucial, yet proves challenging for modern deep networks. Feature visualization (FV) is a powerful tool to decode what information neurons are responding to and hence to better understand the reasoning behind such networks. In particular, in FV we generate human-understandable images that reflect the information detected by neurons of interest. However, current methods often yield unrecognizable visualizations, exhibiting repetitive patterns and visual artifacts that are hard to understand for a human. To address these problems, we propose to guide FV through statistics of real image features combined with measures of relevant network flow to generate prototypical images. Our approach yields human-understandable visualizations that both qualitatively and quantitatively improve over state-of-the-art FVs across various architectures. As such, it can be used to decode which information the network uses, complementing mechanistic circuits that identify where it is encoded. Code is available at: https://github.com/adagorgun/VITAL

Post-hoc importance attribution methods are a popular tool for “explaining” Deep Neural Networks (DNNs) and are inherently based on the assumption that the explanations can be applied independently of how the models were trained. Contrarily, in this work we bring forward empirical evidence that challenges this very notion. Surprisingly, we discover a strong dependency on and demonstrate that the training details of a pre-trained model’s classification layer (<10% of model parameters) play a crucial role, much more than the pre-training scheme itself. This is of high practical relevance: (1) as techniques for pre-training models are becoming increasingly diverse, understanding the interplay between these techniques and attribution methods is critical; (2) it sheds light on an important yet overlooked assumption of post-hoc attribution methods which can drastically impact model explanations and how they are interpreted eventually. With this finding we also present simple yet effective adjustments to the classification layers, that can significantly enhance the quality of model explanations. We validate our findings across several visual pre-training frameworks (fully-supervised, self-supervised, contrastive vision-language training) and analyse how they impact explanations for a wide range of attribution methods on a diverse set of evaluation metrics.

Multiple object tracking in complex scenarios - such as coordinated dance performances, team sports, or dynamic animal groups - presents unique challenges. In these settings, objects frequently move in coordinated patterns, occlude each other, and exhibit long-term dependencies in their trajectories. However, it remains a key open research question on how to model long-range dependencies within tracklets, interdependencies among tracklets, and the associated temporal occlusions. To this end, we introduce Samba, a novel linear-time set-of-sequences model designed to jointly process multiple tracklets by synchronizing the multiple selective state-spaces used to model each tracklet. Samba autoregressively predicts the future track query for each sequence while maintaining synchronized long-term memory representations across tracklets. By integrating Samba into a tracking-by-propagation framework, we propose SambaMOTR, the first tracker effectively addressing the aforementioned issues, including long-range dependencies, tracklet interdependencies, and temporal occlusions. Additionally, we introduce an effective technique for dealing with uncertain observations (MaskObs) and an efficient training recipe to scale SambaMOTR to longer sequences. By modeling long-range dependencies and interactions among tracked objects, SambaMOTR implicitly learns to track objects accurately through occlusions without any hand-crafted heuristics. Our approach significantly surpasses prior state-of-the-art on the DanceTrack, BFT, and SportsMOT datasets.

Transformers have become the predominant architecture in foundation models due to their excellent performance across various domains. However, the substantial cost of scaling these models remains a significant concern. This problem arises primarily from their dependence on a fixed number of parameters within linear projections. When architectural modifications (e.g., channel dimensions) are introduced, the entire model typically requires retraining from scratch. As model sizes continue growing, this strategy results in increasingly high computational costs and becomes unsustainable. To overcome this problem, we introduce Tokenformer, a natively scalable architecture that leverages the attention mechanism not only for computations among input tokens but also for interactions between tokens and model parameters, thereby enhancing architectural flexibility. By treating model parameters as tokens, we replace all the linear projections in Transformers with our token-parameter attention layer, where input tokens act as queries and model parameters as keys and values. This reformulation allows for progressive and efficient scaling without necessitating retraining from scratch. Our model scales from 124M to 1.4B parameters by incrementally adding new key-value parameter pairs, achieving performance comparable to Transformers trained from scratch while greatly reducing training costs. Code and models are available at {\color{red}\url{https://github.com/Haiyang-W/TokenFormer.git}}

Post-hoc attribution methods aim to explain deep learning predictions by highlighting influential input pixels. However, these explanations are highly non-robust: small, imperceptible input perturbations can drastically alter the attribution map while maintaining the same prediction. This vulnerability undermines their trustworthiness and calls for rigorous robustness guarantees of pixel-level attribution scores. We introduce the first certification framework that guarantees pixel-level robustness for any black-box attribution method using randomized smoothing. By sparsifying and smoothing attribution maps, we reformulate the task as a segmentation problem and certify each pixel’s importance against $\ell_2$-bounded perturbations. We further propose three evaluation metrics to assess certified robustness, localization, and faithfulness. An extensive evaluation of 12 attribution methods across 5 ImageNet models shows that our certified attributions are robust, interpretable, and faithful, enabling reliable use in downstream tasks. Our code is at https://github.com/AlaaAnani/certified-attributions.

We introduce Spatial Reasoning Models (SRMs), a framework to perform reasoning over sets of continuous variables via denoising generative models. SRMs infer continuous representations on a set of unobserved variables, given observations on observed variables. Current generative models on spatial domains, such as diffusion and flow matching models, often collapse to hallucination in case of complex distributions. To measure this, we introduce a set of benchmark tasks that test the quality of complex reasoning in generative models and can quantify hallucination. The SRM framework allows to report key findings about importance of sequentialization in generation, the associated order, as well as the sampling strategies during training. It demonstrates, for the first time, that order of generation can successfully be predicted by the denoising network itself. Using these findings, we can increase the accuracy of specific reasoning tasks from $<$1% to $>$50%. Our project website provides additional videos, code, and the benchmark datasets.

Flow-based latent generative models such as Stable Diffusion 3 are able to generate images with remarkable quality, even enabling photorealistic text-to-image generation. Their impressive performance suggests that these models should also constitute powerful priors for inverse imaging problems, but that approach has not yet led to comparable fidelity. There are several key obstacles: (i) the data likelihood term is usually intractable; (ii) learned generative models cannot be directly conditioned on the distorted observations, leading to conflicting objectives between data likelihood and prior; and (iii) the reconstructions can deviate from the observed data. We present FLAIR, a novel, training-free variational framework that leverages flow-based generative models as prior for inverse problems. To that end, we introduce a variational objective for flow matching that is agnostic to the type of degradation, and combine it with deterministic trajectory adjustments to guide the prior towards regions which are more likely under the posterior. To enforce exact consistency with the observed data, we decouple the optimization of the data fidelity and regularization terms. Moreover, we introduce a time-dependent calibration scheme in which the strength of the regularization is modulated according to off-line accuracy estimates. Results on standard imaging benchmarks demonstrate that FLAIR consistently outperforms existing diffusion- and flow-based methods in terms of reconstruction quality and sample diversity. Source code is available at https://inverseflair.github.io/.

Deep networks have shown remarkable performance across a wide range of tasks, yet getting a global concept-level understanding of how they function remains a key challenge. Many post-hoc concept-based approaches have been introduced to understand their workings, yet they are not always faithful to the model. Further, they make restrictive assumptions on the concepts a model learns, such as class-specificity, small spatial extent, or alignment to human expectations. In this work, we put emphasis on the faithfulness of such concept-based explanations and propose a new model with model-inherent mechanistic concept-explanations. Our concepts are shared across classes and, from any layer, their contribution to the logit and their input-visualization can be faithfully traced. We also leverage foundation models to propose a new concept-consistency metric, C$^2$-Score, that can be used to evaluate concept-based methods. We show that, compared to prior work, our concepts are quantitatively more consistent and users find our concepts to be more interpretable, all while retaining competitive ImageNet performance.

Recent advances in multi-modal generative models have enabled significant progress in instruction-based image editing. However, while these models produce visually plausible outputs, their capacity for knowledge-based reasoning editing tasks remains under-explored. In this paper, We introduce KRIS-Bench (Knowledge-based Reasoning in Image-editing Systems Benchmark), a diagnostic benchmark designed to assess models through a cognitively informed lens. Drawing from educational theory, KRIS-Bench categorizes editing tasks across three foundational knowledge types: Factual, Conceptual, and Procedural. Based on this taxonomy, we design 22 representative tasks spanning 7 reasoning dimensions and release 1,267 high-quality annotated editing instances. To support fine-grained evaluation, we propose a comprehensive protocol that incorporates a novel Knowledge Plausibility metric, enhanced by knowledge hints and calibrated through human studies. Empirical results on nine state-of-the-art models reveal significant gaps in reasoning performance, highlighting the need for knowledge-centric benchmarks to advance the development of intelligent image editing systems.

In dynamic 3D environments, the ability to recognize a diverse range of objects without the constraints of predefined categories is indispensable for real-world applications. In response to this need, we introduce OV3D, an innovative framework designed for open-vocabulary 3D semantic segmentation. OV3D leverages the broad open-world knowledge embedded in vision and language foundation models to establish a fine-grained correspondence between 3D points and textual entity descriptions. These entity descriptions are enriched with contextual information, enabling a more open and comprehensive understanding. By seamlessly aligning 3D point features with entity text features, OV3D empowers open-vocabulary recognition in the 3D domain, achieving state-of-the-art open-vocabulary semantic segmentation performance across multiple datasets, including ScanNet, Matterport3D, and nuScenes.

Few-Shot Class-Incremental Learning (FSCIL) intro-duces a paradigm in which the problem space expands with limited data. FSCIL methods inherently face the chal-lenge of catastrophic forgetting as data arrives incremen-tally, making models susceptible to overwriting previously acquired knowledge. Moreover, given the scarcity of la-beled samples available at any given time, models may be prone to overfitting and find it challenging to strike a bal-ance between extensive pretraining and the limited incre-mental data. To address these challenges, we propose the OrCo framework built on two core principles: features' orthogonality in the representation space, and contrastive learning. In particular, we improve the generalization of the embedding space by employing a combination of su-pervised and self-supervised contrastive losses during the pretraining phase. Additionally, we introduce OrCo loss to address challenges arising from data limitations during in-cremental sessions. Through feature space perturbations and orthogonality between classes, the OrCo loss maxi-mizes margins and reserves space for the following incre-mental data. This, in turn, ensures the accommodation of incoming classes in the feature space without compromising previously acquired knowledge. Our experimental results showcase state-of-the-art performance across three bench-mark datasets, including mini-ImageNet, CIFAR100, and CUB datasets. Code is available at: https://github.com/noorahmedds/OrCo.

We present S4Former, a novel approach to training Vision Transformers for Semi-Supervised Semantic Segmentation (S4). At its core, S4Former employs a Vision Transformer within a classic teacher-student framework, and then leverages three novel technical ingredients: PatchShuffle as a parameter-free perturbation technique, Patch-Adaptive Self-Attention (PASA) as a fine-grainedfeature modulation method, and the innovative Negative Class Ranking (NCR) regularization loss. Based on these regu-larization modules aligned with Transformer-specific char-acteristics across the image input, feature, and output di-mensions, S4Former exploits the Transformer's ability to capture and differentiate consistent global contextual information in unlabeled images. Overall, S4 Former not only defines a new state of the art in S4 but also maintains a streamlined and scalable architecture. Being readily compatible with existing frameworks, S4 Former achieves strong improvements (up to 4.9%) on benchmarks like Pascal VOC 2012, COCO, and Cityscapes, with varying numbers of labeled data. The code is at https://github.com/JoyHuYY1412/S4Former.

Lately, there has been growing interest in adapting vision-language models (VLMs) to image and third-person video classification due to their success in zero-shot recog-nition. However, the adaptation of these models to egocentric videos has been largely unexplored. To address this gap, we propose a simple yet effective cross-modal adaptation framework, which we call X-MIC. Using a video adapter, our pipeline learns to align frozen text embeddings to each egocentric video directly in the shared embedding space. Our novel adapter architecture retains and improves generalization of the pre-trained VLMs by disentangling learnable temporal modeling and frozen visual en-coder. This results in an enhanced alignment of text embeddings to each egocentric video, leading to a significant improvement in cross-dataset generalization. We evaluate our approach on the Epic-Kitchens, Ego4D, and EGTEA datasets for fine-grained cross-dataset action generalization, demonstrating the effectiveness of our method.11https://github.com/annusha/xmic

Adversarial training is the de-facto standard for improving robustness against adversarial examples. This usually involves a multi-step adversarial attack applied on each example during training. In this paper, we explore only constructing adversarial examples (AE) on a subset of the training examples. That is, we split the training set in two subsets $A$ and $B$, train models on both ($A\cup B$) but construct AEs only for examples in $A$. Starting with $A$ containing only a single class, we systematically increase the size of $A$ and consider splitting by class and by examples. We observe that: (i) adv. robustness transfers by difficulty and to classes in $B$ that have never been adv. attacked during training, (ii) we observe a tendency for hard examples to provide better robustness transfer than easy examples, yet find this tendency to diminish with increasing complexity of datasets (iii) generating AEs on only $50$% of training data is sufficient to recover most of the baseline AT performance even on ImageNet. We observe similar transfer properties across tasks, where generating AEs on only $30$% of data can recover baseline robustness on the target task. We evaluate our subset analysis on a wide variety of image datasets like CIFAR-10, CIFAR-100, ImageNet-200 and show transfer to SVHN, Oxford-Flowers-102 and Caltech-256. In contrast to conventional practice, our experiments indicate that the utility of computing AEs varies by class and examples and that weighting examples from $A$ higher than $B$ provides high transfer performance. Code is available at [http://github.com/mlosch/SAT](http://github.com/mlosch/SAT).

Certification for machine learning is proving that no adversarial sample can evade a model within a range under certain conditions, a necessity for safety-critical domains. Common certification methods for segmentation use a flat set of fine-grained classes, leading to high abstain rates due to model uncertainty across many classes. We propose a novel, more practical setting, which certifies pixels within a multi-level hierarchy, and adaptively relaxes the certification to a coarser level for unstable components classic methods would abstain from, effectively lowering the abstain rate whilst providing more certified semantically meaningful information. We mathematically formulate the problem setup, introduce an adaptive hierarchical certification algorithm and prove the correctness of its guarantees. Since certified accuracy does not take the loss of information into account for coarser classes, we introduce the Certified Information Gain ($\mathrm{CIG}$) metric, which is proportional to the class granularity level. Our extensive experiments on the datasets Cityscapes, PASCAL-Context, ACDC and COCO-Stuff demonstrate that our adaptive algorithm achieves a higher $\mathrm{CIG}$ and lower abstain rate compared to the current state-of-the-art certification method. Our code can be found here: [https://github.com/AlaaAnani/adaptive-certify](https://github.com/AlaaAnani/adaptive-certify).

B-cos Networks have been shown to be effective for obtaining highly human interpretable explanations of model decisions by architecturally enforcing stronger alignment between inputs and weight. B-cos variants of convolutional networks (CNNs) and vision transformers (ViTs), which primarily replace linear layers with B-cos transformations, perform competitively to their respective standard variants while also yielding explanations that are faithful by design. However, it has so far been necessary to train these models from scratch, which is increasingly infeasible in the era of large, pre-trained foundation models. In this work, inspired by the architectural similarities in standard DNNs and B-cos networks, we propose ‘B-cosification’, a novel approach to transform existing pre-trained models to become inherently interpretable. We perform a thorough study of design choices to perform this conversion, both for convolutional neural networks and vision transformers. We find that B-cosification can yield models that are on par with B-cos models trained from scratch in terms of interpretability, while often outperforming them in terms of classification performance at a fraction of the training cost. Subsequently, we apply B-cosification to a pretrained CLIP model, and show that, even with limited data and compute cost, we obtain a B-cosified version that is highly interpretable and competitive on zero shot performance across a variety of datasets. We release our code and pre-trained model weights at https://github.com/shrebox/B-cosification.

In this work, we introduce Scribbles for All, a label and training data generation algorithm for semantic segmentation trained on scribble labels. Training or fine-tuning semantic segmentation models with weak supervision has become an important topic recently and was subject to significant advances in model quality. In this setting, scribbles are a promising label type to achieve high quality segmentation results while requiring a much lower annotation effort than usual pixel-wise dense semantic segmentation annotations. The main limitation of scribbles as source for weak supervision is the lack of challenging datasets for scribble segmentation, which hinders the development of novel methods and conclusive evaluations. To overcome this limitation, Scribbles for All provides scribble labels for several popular segmentation datasets and provides an algorithm to automatically generate scribble labels for any dataset with dense annotations, paving the way for new insights and model advancements in the field of weakly supervised segmentation. In addition to providing datasets and algorithm, we evaluate state-of-the-art segmentation models on our datasets and show that models trained with our synthetic labels perform competitively with respect to models trained on manual labels. Thus, our datasets enable state-of-the-art research into methods for scribble-labeled semantic segmentation. The datasets, scribble generation algorithm, and baselines are publicly available at https://github.com/wbkit/Scribbles4All.

Not forgetting old class knowledge is a key challenge for class-incremental learning (CIL) when the model continuously adapts to new classes. A common technique to address this is knowledge distillation (KD), which penalizes prediction inconsistencies between old and new models. Such prediction is made with almost new class data, as old class data is extremely scarce due to the strict memory limitation in CIL. In this paper, we take a deep dive into KD losses and find that "using new class data for KD" not only hinders the model adaption (for learning new classes) but also results in low efficiency for preserving old class knowledge. We address this by "using the placebos of old classes for KD", where the placebos are chosen from a free image stream, such as Google Images, in an automatical and economical fashion. To this end, we train an online placebo selection policy to quickly evaluate the quality of streaming images (good or bad placebos) and use only good ones for one-time feed-forward computation of KD. We formulate the policy training process as an online Markov Decision Process (MDP), and introduce an online learning algorithm to solve this MDP problem without causing much computation costs. In experiments, we show that our method 1) is surprisingly effective even when there is no class overlap between placebos and original old class data, 2) does not require any additional supervision or memory budget, and 3) significantly outperforms a number of top-performing CIL methods, in particular when using lower memory budgets for old class exemplars, e.g., five exemplars per class.1

We present a new direction for increasing the interpretability of deep neural networks (DNNs) by promoting weight-input alignment during training. For this, we propose to replace the linear transformations in DNNs by our novel B-cos transformation. As we show, a sequence (network) of such transformations induces a single linear transformation that faithfully summarises the full model computations. Moreover, the B-cos transformation is designed such that the weights align with relevant signals during optimisation. As a result, those induced linear transformations become highly interpretable and highlight task-relevant features. Importantly, the B-cos transformation is designed to be compatible with existing architectures and we show that it can easily be integrated into virtually all of the latest state of the art models for computer vision-e.g. ResNets, DenseNets, ConvNext models, as well as Vision Transformers-by combining the B-cos-based explanations with normalisation and attention layers, all whilst maintaining similar accuracy on ImageNet. Finally, we show that the resulting explanations are of high visual quality and perform well under quantitative interpretability metrics.

Deep neural networks are very successful on many vision tasks, but hard to interpret due to their black box nature. To overcome this, various post-hoc attribution methods have been proposed to identify image regions most influential to the models' decisions. Evaluating such methods is challenging since no ground truth attributions exist. We thus propose three novel evaluation schemes to more reliably measure the faithfulness of those methods, to make comparisons between them more fair, and to make visual inspection more systematic. To address faithfulness, we propose a novel evaluation setting (DiFull) in which we carefully control which parts of the input can influence the output in order to distinguish possible from impossible attributions. To address fairness, we note that different methods are applied at different layers, which skews any comparison, and so evaluate all methods on the same layers (ML-Att) and discuss how this impacts their performance on quantitative metrics. For more systematic visualizations, we propose a scheme (AggAtt) to qualitatively evaluate the methods on complete datasets. We use these evaluation schemes to study strengths and shortcomings of some widely used attribution methods over a wide range of models. Finally, we propose a post-processing smoothing step that significantly improves the performance of some attribution methods, and discuss its applicability.

Motion prediction is crucial for autonomous driving systems to understand complex driving scenarios and make informed decisions. However, this task is challenging due to the diverse behaviors of traffic participants and complex environmental contexts. In this paper, we propose Motion TRansformer (MTR) frameworks to address these challenges. The initial MTR framework utilizes a transformer encoder-decoder structure with learnable intention queries, enabling efficient and accurate prediction of future trajectories. By customizing intention queries for distinct motion modalities, MTR improves multimodal motion prediction while reducing reliance on dense goal candidates. The framework comprises two essential processes: global intention localization, identifying the agent's intent to enhance overall efficiency, and local movement refinement, adaptively refining predicted trajectories for improved accuracy. Moreover, we introduce an advanced MTR++ framework, extending the capability of MTR to simultaneously predict multimodal motion for multiple agents. MTR++ incorporates symmetric context modeling and mutually-guided intention querying modules to facilitate future behavior interaction among multiple agents, resulting in scene-compliant future trajectories. Extensive experimental results demonstrate that the MTR framework achieves state-of-the-art performance on the highly-competitive motion prediction benchmarks, while the MTR++ framework surpasses its precursor, exhibiting enhanced performance and efficiency in predicting accurate multimodal future trajectories for multiple agents.

Negative flips are errors introduced in a classification system when a legacy model is updated. Existing methods to reduce the negative flip rate (NFR) either do so at the expense of overall accuracy by forcing a new model to imitate the old models, or use ensembles, which multiply inference cost prohibitively. We analyze the role of ensembles in reducing NFR and observe that they remove negative flips that are typically not close to the decision boundary, but often exhibit large deviations in the distance among their logits. Based on the observation, we present a method, called Ensemble Logit Difference Inhibition (ELODI), to train a classification system that achieves paragon performance in both error rate and NFR, at the inference cost of a single model. The method distills a homogeneous ensemble to a single student model which is used to update the classification system. ELODI also introduces a generalized distillation objective, Logit Difference Inhibition (LDI), which only penalizes the logit difference of a subset of classes with the highest logit values. On multiple image classification benchmarks, model updates with ELODI demonstrate superior accuracy retention and NFR reduction.

Class-incremental learning (CIL) aims to train a classification model while the number of classes increases phase-by-phase. An inherent challenge of CIL is the stability-plasticity tradeoff, i.e., CIL models should keep stable to retain old knowledge and keep plastic to absorb new knowledge. However, none of the existing CIL models can achieve the optimal tradeoff in different data-receiving settings—where typically the training-from-half (TFH) setting needs more stability, but the training-from-scratch (TFS) needs more plasticity. To this end, we design an online learning method that can adaptively optimize the tradeoff without knowing the setting as a priori. Specifically, we first introduce the key hyperparameters that influence the tradeoff, e.g., knowledge distillation (KD) loss weights, learning rates, and classifier types. Then, we formulate the hyperparameter optimization process as an online Markov Decision Process (MDP) problem and propose a specific algorithm to solve it. We apply local estimated rewards and a classic bandit algorithm Exp3 to address the issues when applying online MDP methods to the CIL protocol. Our method consistently improves top-performing CIL methods in both TFH and TFS settings, e.g., boosting the average accuracy of TFH and TFS by 2.2 percentage points on ImageNet-Full, compared to the state-of-the-art. Code is provided at https://class-il.mpi-inf.mpg.de/online/

Incremental object detection (IOD) aims to train an object detector in phases, each with annotations for new object categories. As other incremental settings, IOD is subject to catastrophic forgetting, which is often addressed by techniques such as knowledge distillation (KD) and exem-plar replay (ER). However, KD and ER do not work well if applied directly to state-of-the-art transformer-based object detectors such as Deformable DETR [59] and UP-DETR [9]. In this paper, we solve these issues by proposing a ContinuaL DEtection TRansformer (CL-DETR), a new method for transformer-based IOD which enables effective usage of KD and ER in this context. First, we introduce a Detector Knowledge Distillation (DKD) loss, focusing on the most informative and reliable predictions from old versions of the model, ignoring redundant background predictions, and ensuring compatibility with the available ground-truth labels. We also improve ER by proposing a calibration strategy to preserve the label distribution of the training set, therefore better matching training and testing statistics. We conduct extensive experiments on COCO 2017 and demonstrate that CL-DETR achieves state-of-the-art results in the IOD setting.11Code: https://1yy.mpi-inf.mpg.de/CL-DETR/

There has been a lot of effort in improving the performance of unsupervised domain adaptation for semantic segmentation task, however, there is still a huge gap in performance when compared with supervised learning. In this work, we propose a common framework to use different weak labels, e.g., image, point and coarse labels from the target domain to reduce this performance gap. Specifically, we propose to learn better prototypes that are representative class features by exploiting these weak labels. We use these improved prototypes for the contrastive alignment of class features. In particular, we perform two different feature alignments: first, we align pixel features with proto-types within each domain and second, we align pixel features from the source to prototype of target domain in an asymmetric way. This asymmetric alignment is beneficial as it preserves the target features during training, which is essential when weak labels are available from the target domain. Our experiments on various benchmarks show that our framework achieves significant improvement compared to existing works and can reduce the performance gap with supervised learning. Code will be available at https://github.com/anurag-198/WDASS.

Current semantic segmentation models have achieved great success under the independent and identically distributed (i.i.d.) condition. However, in real-world applications, test data might come from a different domain than training data. Therefore, it is important to improve model robustness against domain differences. This work studies semantic segmentation under the domain generalization setting, where a model is trained only on the source domain and tested on the unseen target domain. Existing works show that Vision Transformers are more robust than CNNs and show that this is related to the visual grouping property of self-attention. In this work, we propose a novel hierarchical grouping transformer (HGFormer) to explicitly group pixels to form part-level masks and then whole-level masks. The masks at different scales aim to segment out both parts and a whole of classes. HGFormer combines mask classification results at both scales for class label prediction. We assemble multiple interesting cross-domain settings by using seven public semantic segmentation datasets. Experiments show that HGFormer yields more robust semantic segmentation results than per-pixel classification methods and flat-grouping transformers, and outperforms previous methods significantly. Code will be available at https://github.com/dingjianswl0l/HGFormer.

Despite their success, vision transformers still remain vulnerable to image corruptions, such as noise or blur. Indeed, we find that the vulnerability mainly stems from the unstable self-attention mechanism, which is inherently built upon patch-based inputs and often becomes overly sensitive to the corruptions across patches. For example, when we only occlude a small number of patches with random noise (e.g., 10%), these patch corruptions would lead to severe accuracy drops and greatly distract intermediate attention layers. To address this, we propose a new training method that improves the robustness of transformers from a new perspective - reducing sensitivity to patch corruptions (RSPC). Specifically, we first identify and occlude/corrupt the most vulnerable patches and then explicitly reduce sensitivity to them by aligning the intermediate features between clean and corrupted examples. We highlight that the construction of patch corruptions is learned adversarially to the following feature alignment process, which is particularly effective and essentially different from existing methods. In experiments, our RSPC greatly improves the stability of attention layers and consistently yields better robustness on various benchmarks, including CIFAR-10/100-C, ImageNet-A, ImageNet-C, and ImageNet-P.

We propose an approach to estimate the number of samples required for a model to reach a target performance. We find that the power law, the de facto principle to estimate model performance, leads to a large error when using a small dataset (e.g., 5 samples per class) for extrapolation. This is because the log-performance error against the log-dataset size follows a nonlinear progression in the few-shot regime followed by a linear progression in the high-shot regime. We introduce a novel piecewise power law (PPL) that handles the two data regimes differently. To estimate the parameters of the PPL, we introduce a random forest regressor trained via meta learning that generalizes across classification/detection tasks, ResNet/ViT based architectures, and random/pre-trained initializations. The PPL improves the performance estimation on average by 37% across 16 classification and 33% across 10 detection datasets, compared to the power law. We further extend the PPL to provide a confidence bound and use it to limit the prediction horizon that reduces over-estimation of data by 76% on classification and 91% on detection datasets.

Masked signal modeling has greatly advanced self-supervised pre-training for language and 2D images. However, it is still not fully explored in 3D scene understanding. Thus, this paper introduces Masked Shape Prediction (MSP), a new framework to conduct masked signal modeling in 3D scenes. MSP uses the essential 3D semantic cue, i.e., geometric shape, as the prediction target for masked points. The context-enhanced shape target consisting of explicit shape context and implicit deep shape feature is proposed to facilitate exploiting contextual cues in shape prediction. Meanwhile, the pre-training architecture in MSP is carefully designed to alleviate the masked shape leakage from point coordinates. Experiments on multiple 3D understanding tasks on both indoor and outdoor datasets demon-strate the effectiveness of MSP in learning good feature representations to consistently boost downstream performance.

Exemplar-based class-incremental learning (CIL) [36] finetunes the model with all samples of new classes but few-shot exemplars of old classes in each incremental phase, where the “few-shot” abides by the limited memory budget. In this paper, we break this “few-shot” limit based on a simple yet surprisingly effective idea: compressing exemplars by downsampling non-discriminative pixels and saving “many-shot” compressed exemplars in the memory. Without needing any manual annotation, we achieve this compression by generating 0–1 masks on discriminative pixels from class activation maps (CAM) [49]. We propose an adaptive mask generation model called class-incremental masking (CIM) to explicitly resolve two difficulties of using CAM: 1) transforming the heatmaps of CAM to 0–1 masks with an arbitrary threshold leads to a trade-off between the coverage on discriminative pixels and the quantity of exemplars, as the total memory is fixed; and 2) optimal thresholds vary for different object classes, which is particularly obvious in the dynamic environment of CIL. We optimize the CIM model alternatively with the conventional CIL model through a bilevel optimization problem [40]. We conduct extensive experiments on high-resolution CIL benchmarks including Food-101, ImageNet-100, and ImageNet-1000, and show that using the compressed exemplars by CIM can achieve a new state-of-the-art CIL accuracy, e.g., 4.8 percentage points higher than FOSTER [42] on 10-Phase ImageNet-1000. Our code is available at https://github.com/xfflzlICIM-CIL.

Designing an efficient yet deployment-friendly 3D backbone to handle sparse point clouds is a fundamental problem in 3D perception. Compared with the customized sparse convolution, the attention mechanism in Transformers is more appropriate for flexibly modeling long-range relationships and is easier to be deployed in real-world applications. However, due to the sparse characteristics of point clouds, it is non-trivial to apply a standard transformer on sparse points. In this paper, we present Dynamic Sparse Voxel Transformer (DSVT), a single-stride window-based voxel Transformer backbone for outdoor 3D perception. In order to efficiently process sparse points in parallel, we propose Dynamic Sparse Window Attention, which partitions a series of local regions in each window according to its sparsity and then computes the features of all regions in a fully parallel manner. To allow the cross-set connection, we design a rotated set partitioning strategy that alternates between two partitioning configurations in consecutive self-attention layers. To support effective downsampling and better encode geometric information, we also propose an attention-style 3D pooling module on sparse points, which is powerful and deployment-friendly without utilizing any customized CUDA operations. Our model achieves state-of-the-art performance with a broad range of 3D perception tasks. More importantly, DSVT can be easily deployed by TensorRT with real-time inference speed (27Hz). Code will be available at https://github.com/Haiyang-W/DSVT.

In this paper, we show that recent advances in video representation learning and pre-trained vision-language models allow for substantial improvements in self-supervised video object localization. We propose a method that first localizes objects in videos via a slot attention approach and then assigns text to the obtained slots. The latter is achieved by an unsupervised way to read localized semantic information from the pre-trained CLIP model. The resulting video object localization is entirely unsupervised apart from the implicit annotation contained in CLIP, and it is effectively the first unsupervised approach that yields good results on regular video benchmarks.

Semi-supervised learning (SSL) methods effectively leverage unlabeled data to improve model generalization. However, SSL models often underperform in open-set scenarios, where unlabeled data contain outliers from novel categories that do not appear in the labeled set. In this paper, we study the challenging and realistic open-set SSL setting, where the goal is to both correctly classify inliers and to detect outliers. Intuitively, the inlier classifier should be trained on inlier data only. However, we find that inlier classification performance can be largely improved by incorporating high-confidence pseudo-labeled data, regardless of whether they are inliers or outliers. Also, we propose to utilize non-linear transformations to separate the features used for inlier classification and outlier detection in the multi-task learning framework, preventing adverse effects between them. Additionally, we introduce pseudo-negative mining, which further boosts outlier detection performance. The three ingredients lead to what we call Simple but Strong Baseline (SSB) for open-set SSL. In experiments, SSB greatly improves both inlier classification and outlier detection performance, outperforming existing methods by a large margin. Our code will be released at https://github.com/YUE-FAN/SSB.

Despite the success of vision transformers (ViTs), they still suffer from significant drops in accuracy in the presence of common corruptions, such as noise or blur. Interestingly, we observe that the attention mechanism of ViTs tends to rely on few important tokens, a phenomenon we call token overfocusing. More critically, these tokens are not robust to corruptions, often leading to highly diverging attention patterns. In this paper, we intend to alleviate this overfocusing issue and make attention more stable through two general techniques: First, our Token-aware Average Pooling (TAP) module encourages the local neighborhood of each token to take part in the attention mechanism. Specifically, TAP learns average pooling schemes for each token such that the information of potentially important tokens in the neighborhood can adaptively be taken into account. Second, we force the output tokens to aggregate information from a diverse set of input tokens rather than focusing on just a few by using our Attention Diversification Loss (ADL). We achieve this by penalizing high cosine similarity between the attention vectors of different tokens. In experiments, we apply our methods to a wide range of transformer architectures and improve robustness significantly. For example, we improve corruption robustness on ImageNet-C by 2.4% while improving accuracy by 0.4% based on state-of-the-art robust architecture FAN. Also, when fine-tuning on semantic segmentation tasks, we improve robustness on CityScapes-C by 2.4% and ACDC by 3.0%. Our code is available at https://github.com/guoyongcs/TAPADL.

Despite being highly performant, deep neural networks might base their decisions on features that spuriously correlate with the provided labels, thus hurting generalization. To mitigate this, ‘model guidance’ has recently gained popularity, i.e. the idea of regularizing the models’ explanations to ensure that they are "right for the right reasons" [49]. While various techniques to achieve such model guidance have been proposed, experimental validation of these approaches has thus far been limited to relatively simple and / or synthetic datasets. To better understand the effectiveness of the various design choices that have been explored in the context of model guidance, in this work we conduct an in-depth evaluation across various loss functions, attribution methods, models, and ‘guidance depths’ on the PASCAL VOC 2007 and MS COCO 2014 datasets. As annotation costs for model guidance can limit its applicability, we also place a particular focus on efficiency. Specifically, we guide the models via bounding box annotations, which are much cheaper to obtain than the commonly used segmentation masks, and evaluate the robustness of model guidance under limited (e.g. with only 1% of annotated images) or overly coarse annotations. Further, we propose using the EPG score as an additional evaluation metric and loss function (‘Energy loss’). We show that optimizing for the Energy loss leads to models that exhibit a distinct focus on object-specific features, despite only using bounding box annotations that also include background regions. Lastly, we show that such model guidance can improve generalization under distribution shifts. Code available at: https://github.com/sukrutrao/Model-Guidance

Multiple object tracking (MOT) is a fundamental component of perception systems for autonomous driving, and its robustness to unseen conditions is a requirement to avoid life-critical failures. Despite the urge of safety in driving systems, no solution to the MOT adaptation problem to domain shift in test-time conditions has ever been proposed. However, the nature of a MOT system is manifold - requiring object detection and instance association - and adapting all its components is non-trivial. In this paper, we analyze the effect of domain shift on appearance-based trackers, and introduce DARTH, a holistic test-time adaptation framework for MOT. We propose a detection consistency formulation to adapt object detection in a self-supervised fashion, while adapting the instance appearance representations via our novel patch contrastive loss. We evaluate our method on a variety of domain shifts - including sim-to-real, outdoor-to-indoor, indoor-to-outdoor - and substantially improve the source model performance on all metrics. Project page: https://www.vis.xyz/pub/darth.

Large-scale noisy web image-text datasets have been proven to be efficient for learning robust vision-language models. However, to transfer them to the task of video retrieval, models still need to be fine-tuned on hand-curated paired text-video data to adapt to the diverse styles of video descriptions. To address this problem without the need for hand-annotated pairs, we propose a new setting, text-video retrieval with uncurated & unpaired data, that uses only text queries together with uncurated web videos during training without any paired text-video data. To this end, we propose an approach, In-Style, that learns the style of the text queries and transfers it to uncurated web videos. Moreover, to improve generalization, we show that one model can be trained with multiple text styles. To this end, we introduce a multi-style contrastive training procedure, that improves the generalizability over several datasets simultaneously. We evaluate our model on retrieval performance over multiple datasets to demonstrate the advantages of our style transfer framework on the new task of uncurated & unpaired text-video retrieval and improve state-of-the-art performance on zero-shot text-video retrieval.1

Contrastive learning has become an important tool in learning representations from unlabeled data mainly relying on the idea of minimizing distance between positive data pairs, e.g., views from the same images, and maximizing distance between negative data pairs, e.g., views from different images. This paper proposes a new variation of the contrastive learning objective, Group Ordering Constraints (GroCo), that leverages the idea of sorting the distances of positive and negative pairs and computing the respective loss based on how many positive pairs have a larger distance than the negative pairs, and thus are not ordered correctly. To this end, the GroCo loss is based on differentiable sorting networks, which enable training with sorting supervision by matching a differentiable permutation matrix, which is produced by sorting a given set of scores, to a respective ground truth permutation matrix. Applying this idea to groupwise pre-ordered inputs of multiple positive and negative pairs allows introducing the GroCo loss with implicit emphasis on strong positives and negatives, leading to better optimization of the local neighborhood. We evaluate the proposed formulation on various self-supervised learning benchmarks and show that it not only leads to improved results compared to vanilla contrastive learning but also shows competitive performance to comparable methods in linear probing and outperforms current methods in k-NN performance.1

Jointly processing information from multiple sensors is crucial to achieving accurate and robust perception for reliable autonomous driving systems. However, current 3D perception research follows a modality-specific paradigm, leading to additional computation overheads and inefficient collaboration between different sensor data. In this paper, we present an efficient multi-modal backbone for outdoor 3D perception named UniTR, which processes a variety of modalities with unified modeling and shared parameters. Unlike previous works, UniTR introduces a modality-agnostic transformer encoder to handle these view-discrepant sensor data for parallel modal-wise representation learning and automatic cross-modal interaction without additional fusion steps. More importantly, to make full use of these complementary sensor types, we present a novel multi-modal integration strategy by both considering semantic-abundant 2D perspective and geometry-aware 3D sparse neighborhood relations. UniTR is also a fundamentally task-agnostic backbone that naturally supports different 3D perception tasks. It sets a new state-of-the-art performance on the nuScenes benchmark, achieving +1.1 NDS higher for 3D object detection and +12.0 higher mIoU for BEV map segmentation with lower inference latency. Code will be available at https://github.com/Haiyang-W/UniTR.

Existing neural field representations for 3D object reconstruction either (1) utilize object-level representations, but suffer from low-quality details due to conditioning on a global latent code, or (2) are able to perfectly reconstruct the observations, but fail to utilize object-level prior knowledge to infer unobserved regions. We present SimNP, a method to learn category-level self-similarities, which combines the advantages of both worlds by connecting neural point radiance fields with a category-level self-similarity representation. Our contribution is two-fold. (1) We design the first neural point representation on a category level by utilizing the concept of coherent point clouds. The resulting neural point radiance fields store a high level of detail for locally supported object regions. (2) We learn how information is shared between neural points in an unconstrained and unsupervised fashion, which allows to derive unobserved regions of an object during the reconstruction process from given observations. We show that SimNP is able to outperform previous methods in reconstructing symmetric unseen object regions, surpassing methods that build upon category-level or pixel-aligned radiance fields, while providing semantic correspondences between instances.

Unsupervised object-centric learning methods allow the partitioning of scenes into entities without additional localization information and are excellent candidates for reducing the annotation burden of multiple-object tracking (MOT) pipelines. Unfortunately, they lack two key properties: objects are often split into parts and are not consistently tracked over time. In fact, state-of-the-art models achieve pixel-level accuracy and temporal consistency by relying on supervised object detection with additional ID labels for the association through time. This paper proposes a video object-centric model for MOT. It consists of an index-merge module that adapts the object-centric slots into detection outputs and an object memory module that builds complete object prototypes to handle occlusions. Benefited from object-centric learning, we only require sparse detection labels (0%-6.25%) for object localization and feature binding. Relying on our self-supervised Expectation-Maximization-inspired loss for object association, our approach requires no ID labels. Our experiments significantly narrow the gap between the existing object-centric model and the fully supervised state-of-the-art and outperform several unsupervised trackers. Code is available at https://github.com/amazon-science/object-centric-multiple-object-tracking.

Test-time domain adaptation, i.e. adapting source-pretrained models to the test data on-the-fly in a source-free, unsupervised manner, is a highly practical yet very challenging task. Due to the domain gap between source and target data, inference quality on the target domain can drop drastically especially in terms of absolute scale of depth. In addition, unsupervised adaptation can degrade the model performance due to inaccurate pseudo labels. Furthermore, the model can suffer from catastrophic forgetting when errors are accumulated over time. We propose a test-time domain adaptation framework for monocular depth estimation which achieves both stability and adaptation performance by benefiting from both self-training of the supervised branch and pseudo labels from self-supervised branch, and is able to tackle the above problems: our scale alignment scheme aligns the input features between source and target data, correcting the absolute scale inference on the target domain; with pseudo label consistency check, we select confident pixels thus improve pseudo label quality; regularisation and self-training schemes are applied to help avoid catastrophic forgetting. Without requirement of further supervisions on the target domain, our method adapts the source-trained models to the test data with significant improvements over the direct inference results, providing scale-aware depth map outputs that outperform the state-of-the-arts. Code is available at https://github.com/Malefikus/ada-depth.

For best performance, today’s semantic segmentation methods use large and carefully labeled datasets, requiring expensive annotation budgets. In this work, we show that coarse annotation is a low-cost but highly effective alternative for training semantic segmentation models. Considering the urban scene segmentation scenario, we lever-age cheap coarse annotations for real-world captured data, as well as synthetic data to train our model and show competitive performance compared with finely annotated real-world data. Specifically, we propose a coarse-to-fine self-training framework that generates pseudo labels for unlabeled regions of the coarsely annotated data, using synthetic data to improve predictions around the boundaries between semantic classes, and using cross-domain data augmentation to increase diversity. Our extensive experimental results on Cityscapes and BDD100k datasets demonstrate that our method achieves a significantly better performance vs annotation cost tradeoff, yielding a comparable performance to fully annotated data with only a small fraction of the annotation budget. Also, when used as pre-training, our framework performs better compared to the standard fully supervised setting.

Despite their recent success, deep neural networks continue to perform poorly when they encounter distribution shifts at test time. Many recently proposed approaches try to counter this by aligning the model to the new distribution prior to inference. With no labels available this requires unsupervised objectives to adapt the model on the observed test data. In this paper, we propose Test-Time Self-Training (TeST): a technique that takes as input a model trained on some source data and a novel data distribution at test time, and learns invariant and robust representations using a student-teacher framework. We find that models adapted using TeST significantly improve over baseline test-time adaptation algorithms. TeST achieves competitive performance to modern domain adaptation algorithms [4], [43], while having access to 5-10x less data at time of adaption. We thoroughly evaluate a variety of baselines on two tasks: object detection and image segmentation and find that models adapted with TeST. We find that TeST sets the new state-of-the art for test-time domain adaptation algorithms.

We introduce a new family of neural network models called Convolutional Dynamic Alignment Networks (CoDA Nets), which are performant classifiers with a high degree of inherent interpretability. Their core building blocks are Dynamic Alignment Units (DAUs), which are optimised to transform their inputs with dynamically computed weight vectors that align with task-relevant patterns. As a result, CoDA Nets model the classification prediction through a series of input-dependent linear transformations, allowing for linear decomposition of the output into individual input contributions. Given the alignment of the DAUs, the resulting contribution maps align with discriminative input patterns. These model-inherent decompositions are of high visual quality and outperform existing attribution methods under quantitative metrics. Further, CoDA Nets constitute performant classifiers, achieving on par results to ResNet and VGG models on e.g. CIFAR-10 and TinyImagenet. Lastly, CoDA Nets can be combined with conventional neural network models to yield powerful classifiers that more easily scale to complex datasets such as Imagenet whilst exhibiting an increased interpretable depth, i.e., the output can be explained well in terms of contributions from intermediate layers within the network.

Deep neural network (DNN) accelerators received considerable attention in recent years due to the potential to save energy compared to mainstream hardware. Low-voltage operation of DNN accelerators allows to further reduce energy consumption, however, causes bit-level failures in the memory storing the quantized weights. Furthermore, DNN accelerators are vulnerable to adversarial attacks on voltage controllers or individual bits. In this paper, we show that a combination of robust fixed-point quantization, weight clipping, as well as random bit error training (RandBET) or adversarial bit error training (AdvBET) improves robustness against random or adversarial bit errors in quantized DNN weights significantly. This leads not only to high energy savings for low-voltage operation as well as low-precision quantization, but also improves security of DNN accelerators. In contrast to related work, our approach generalizes across operating voltages and accelerators and does not require hardware changes. Moreover, we present a novel adversarial bit error attack and are able to obtain robustness against both targeted and untargeted bit-level attacks. Without losing more than 0.8%/2% in test accuracy, we can reduce energy consumption on CIFAR10by 20%/30% for 8/4-bit quantization. Allowing up to 320 adversarial bit errors, we reduce test error from above 90% (chance level) to 26.22%.

Abstract Current work on image-based story generation suffers from the fact that the existing image sequence collections do not have coherent plots behind them. We improve visual story generation by producing a new image-grounded dataset, Visual Writing Prompts (VWP). VWP contains almost 2K selected sequences of movie shots, each including 5-10 images. The image sequences are aligned with a total of 12K stories which were collected via crowdsourcing given the image sequences and a set of grounded characters from the corresponding image sequence. Our new image sequence collection and filtering process has allowed us to obtain stories that are more coherent, diverse, and visually grounded compared to previous work. We also propose a character-based story generation model driven by coherence as a strong baseline. Evaluations show that our generated stories are more coherent, visually grounded, and diverse than stories generated with the current state-of-the-art model. Our code, image features, annotations and collected stories are available at https://vwprompt.github.io/.

3D human motion capture from monocular RGB images respecting interactions of a subject with complex and possibly deformable environments is a very challenging, illposed and under-explored problem. Existing methods address it only weakly and do not model possible surface deformations often occurring when humans interact with scene surfaces. In contrast, this paper proposes MoCapDe-form, i.e., a new framework for monocular 3D human motion capture that is the first to explicitly model non-rigid deformations of a 3D scene for improved 3D human pose estimation and deformable environment reconstruction. Mo-CapDeform accepts a monocular RGB video and a 3D scene mesh aligned in the camera space. It first localises a subject in the input monocular video along with dense contact labels using a new raycasting based strategy. Next, our human-environment interaction constraints are leveraged to jointly optimise global 3D human poses and non-rigid surface deformations. MoCapDeform achieves superior accuracy than competing methods on several datasets, including our newly recorded one with deforming background scenes.

We address the task of estimating the 3D shapes of novel shape classes from a single RGB image. Prior works are either limited to reconstructing known training classes or are unable to reconstruct high-quality shapes. To solve those issues, we propose Generalizing Implicit Networks (GIN) which decomposes 3D reconstruction into 1.) front-back depth estimation followed by differentiable depth voxelization, and 2.) implicit shape completion with 3D features. The key insight is that the depth estimation network learns local class-agnostic shape priors, allowing us to generalize to novel classes, while our implicit shape completion network is able to predict accurate shapes with rich details by learning implicit surfaces in 3D voxel space. We conduct extensive experiments on a large-scale benchmark using 55 classes of ShapeNet and real images of Pix3D. We qualitatively and quantitatively show that the proposed GIN significantly outperforms the state of the art on both seen and novel shape classes for single-image 3D reconstruction. We also illustrate that our GIN can be further improved by using only few-shot depth supervision from novel classes.

Most learning methods for 3D data suffer significant performance drops when the data is not carefully aligned to a canonical orientation. Aligning real world 3D data collected from different sources is non-trivial and requires manual intervention. In this paper, we propose the Adjoint Rigid Transform (ART) Network, a neural module which can be integrated with a variety of 3D networks to significantly boost their performance. ART learns to rotate input shapes to a learned canonical orientation, which is crucial for a lot of tasks such as shape reconstruction, interpolation, non-rigid registration, and latent disentanglement. ART achieves this with self-supervision and a rotation equivariance constraint on predicted rotations. With only self-supervision, ART facilitates learning a canonical orientation for both rigid and nonrigid shapes, which leads to a notable boost in performance of aforementioned tasks. Our code and model are available at [2].

Video-based person re-identification~(re-ID) is an important technique in visual surveillance systems which aims to match video snippets of people captured by different cameras. Existing methods are mostly based on convolutional neural networks~(CNNs), whose building blocks either process local neighbor pixels at a time, or, when 3D convolutions are used to model temporal information, suffer from the misalignment problem caused by person movement. In this paper, we propose to overcome the limitations of normal convolutions with a human-oriented graph method. Specifically, features located at person joint keypoints are extracted and connected as a spatial-temporal graph. These keypoint features are then updated by message passing from their connected nodes with a graph convolutional network~(GCN). During training, the GCN can be attached to any CNN-based person re-ID model to assist representation learning on feature maps, whilst it can be dropped after training for better inference speed. Our method brings significant improvements over the CNN-based baseline model on the MARS dataset with generated person keypoints and a newly annotated dataset: PoseTrackReID. It also defines a new state-of-the-art method in terms of top-1 accuracy and mean average precision in comparison to prior works.

We present a new direction for increasing the interpretability of deep neural networks (DNNs) by promoting weight-input alignment during training. For this, we propose to replace the linear transforms in DNNs by our B-cos transform. As we show, a sequence (network) of such transforms induces a single linear transform that faithfully summarises the full model computations. Moreover, the B-cos transform introduces alignment pressure on the weights during optimisation. As a result, those induced linear transforms become highly interpretable and align with task-relevant features. Importantly, the B-cos transform is designed to be compatible with existing architectures and we show that it can easily be integrated into common models such as VGGs, ResNets, InceptionNets, and DenseNets, whilst maintaining similar performance on ImageNet. The resulting explanations are of high visual quality and perform well under quantitative metrics for interpretability. Code available at github.com/moboehle/B-cos.

Current research evaluates person search, multi-object tracking and multi-person pose estimation as separate tasks and on different datasets although these tasks are very akin to each other and comprise similar sub-tasks, e.g. person detection or appearance-based association of detected persons. Consequently, approaches on these respective tasks are eligible to complement each other. Therefore, we introduce PoseTrack21, a large-scale dataset for person search, multi-object tracking and multi-person pose tracking in real-world scenarios with a high diversity of poses. The dataset provides rich annotations like human pose annotations including annotations of joint occlusions, bounding box annotations even for small persons, and person-ids within and across video sequences. The dataset allows to evaluate multi-object tracking and multi-person pose tracking jointly with person re-identification or exploit structural knowledge of human poses to improve person search and tracking, particularly in the context of severe occlusions. With PoseTrack21, we want to encourage researchers to work on joint approaches that perform reasonably well on all three tasks.

Recently, crowd density estimation has received increasing attention. The main challenge for this task is to achieve high-quality manual annotations on a large amount of training data. To avoid reliance on such annotations, previous works apply unsupervised domain adaptation (UDA) techniques by transferring knowledge learned from easily accessible synthetic data to real-world datasets. However, current state-of-the-art methods either rely on external data for training an auxiliary task or apply an expensive coarse-to-fine estimation. In this work, we aim to develop a new adversarial learning based method, which is simple and efficient to apply. To reduce the domain gap between the synthetic and real data, we design a bi-level alignment framework (BLA) consisting of (1) task-driven data alignment and (2) fine-grained feature alignment. In contrast to previous domain augmentation methods, we introduce AutoML to search for an optimal transform on source, which well serves for the downstream task. On the other hand, we do fine-grained alignment for foreground and background separately to alleviate the alignment difficulty. We evaluate our approach on five real-world crowd counting benchmarks, where we outperform existing approaches by a large margin. Also, our approach is simple, easy to implement and efficient to apply. The code is publicly available at https://github.com/Yankeegsj/BLA.

Deep neural networks are very successful on many vision tasks, but hard to interpret due to their black box nature. To overcome this, various post-hoc attribution methods have been proposed to identify image regions most influential to the models' decisions. Evaluating such methods is challenging since no ground truth attributions exist. We thus propose three novel evaluation schemes to more reliably measure the faithfulness of those methods, to make comparisons between them more fair, and to make visual inspection more systematic. To address faithfulness, we propose a novel evaluation setting (DiFull) in which we carefully control which parts of the input can influence the output in order to distinguish possible from impossible attributions. To address fairness, we note that different methods are applied at different layers, which skews any comparison, and so evaluate all methods on the same layers (ML-Att) and discuss how this impacts their performance on quantitative metrics. For more systematic visualizations, we propose a scheme (AggAtt) to qualitatively evaluate the methods on complete datasets. We use these evaluation schemes to study strengths and shortcomings of some widely used attribution methods. Finally, we propose a post-processing smoothing step that significantly improves the performance of some attribution methods, and discuss its applicability.

Adapting to a continuously evolving environment is a safety-critical challenge inevitably faced by all autonomous-driving systems. Existing image- and video-based driving datasets, however, fall short of capturing the mutable nature of the real world. In this paper, we introduce the largest multi-task synthetic dataset for autonomous driving, SHIFT. It presents discrete and continuous shifts in cloudiness, rain and fog intensity, time of day, and vehicle and pedestrian density. Featuring a comprehensive sensor suite and annotations for several mainstream perception tasks, SHIFT allows to investigate how a perception systems' performance degrades at increasing levels of domain shift, fostering the development of continuous adaptation strategies to mitigate this problem and assessing the robustness and generality of a model. Our dataset and benchmark toolkit are publicly available at www.vis.xyz/shift.

We consider the problem of omni-supervised object detection, which can use unlabeled, fully labeled and weakly labeled annotations, such as image tags, counts, points, etc., for object detection. This is enabled by a unified architecture, Omni-DETR, based on the recent progress on student-teacher framework and end-to-end transformer based object detection. Under this unified architecture, different types of weak labels can be leveraged to generate accurate pseudo labels, by a bipartite matching based filtering mechanism, for the model to learn. In the experiments, Omni-DETR has achieved state-of-the-art results on multiple datasets and settings. And we have found that weak annotations can help to improve detection performance and a mixture of them can achieve a better trade-off between annotation cost and accuracy than the standard complete annotation. These findings could encourage larger object detection datasets with mixture annotations. The code is available at https://github.com/amazon-research/omni-detr.

As a fundamental problem in computer vision, 3D object detection is experiencing rapid growth. To extract the point-wise features from the irregularly and sparsely distributed points, previous methods usually take a feature grouping module to aggregate the point features to an object candidate. However, these methods have not yet leveraged the surface geometry of foreground objects to enhance grouping and 3D box generation. In this paper, we propose the RBGNet framework, a voting-based 3D detector for accurate 3D object detection from point clouds. In order to learn better representations of object shape to enhance cluster features for predicting 3D boxes, we propose a ray-based feature grouping module, which aggregates the point-wise features on object surfaces using a group of determined rays uniformly emitted from cluster centers. Considering the fact that foreground points are more meaningful for box estimation, we design a novel foreground biased sampling strategy in downsample process to sample more points on object surfaces and further boost the detection performance. Our model achieves state-of-the-art 3D detection performance on ScanNet V2 and SUN RGB-D with remarkable performance gains. Code will be available at https://github.com/Haiyang-W/RBGNet.

Human-annotated attributes serve as powerful semantic embeddings in zero-shot learning. However, their annotation process is labor-intensive and needs expert supervision. Current unsupervised semantic embeddings, i.e., word embeddings, enable knowledge transfer between classes. However, word embeddings do not always reflect visual similarities and result in inferior zero-shot performance. We propose to discover semantic embeddings containing discriminative visual properties for zero-shot learning, without requiring any human annotation. Our model visually divides a set of images from seen classes into clusters of local image regions according to their visual similarity, and further imposes their class discrimination and semantic relatedness. To associate these clusters with previously unseen classes, we use external knowledge, e.g., word embeddings and propose a novel class relation discovery module. Through quantitative and qualitative evaluation, we demonstrate that our model discovers semantic embeddings that model the visual properties of both seen and unseen classes. Furthermore, we demonstrate on three benchmarks that our visually-grounded semantic embeddings further improve performance over word embeddings across various ZSL models by a large margin. Code is available at https://github.com/wenjiaXu/VGSE

3D point cloud understanding is an important component in autonomous driving and robotics. In this paper, we present a novel Embedding-Querying paradigm (EQ-Paradigm) for 3D understanding tasks including detection, segmentation and classification. EQ-Paradigm is a unified paradigm that enables combination of existing 3D back-bone architectures with different task heads. Under the EQ-Paradigm, the input is first encoded in the embedding stage with an arbitrary feature extraction architecture, which is independent of tasks and heads. Then, the querying stage enables the encoded features for diverse task heads. This is achieved by introducing an intermediate representation, i.e., Q-representation, in the querying stage to bridge the embedding stage and task heads. We design a novel Q-Net as the querying stage network. Extensive experimental results on various 3D tasks show that EQ-Paradigm in tandem with Q-Net is a general and effective pipeline, which enables flexible collaboration of backbones and heads. It further boosts performance of state-of-the-art methods.

Predicting multimodal future behavior of traffic participants is essential for robotic vehicles to make safe decisions. Existing works explore to directly predict future trajectories based on latent features or utilize dense goal candidates to identify agent's destinations, where the former strategy converges slowly since all motion modes are derived from the same feature while the latter strategy has efficiency issue since its performance highly relies on the density of goal candidates. In this paper, we propose the Motion TRansformer (MTR) framework that models motion prediction as the joint optimization of global intention localization and local movement refinement. Instead of using goal candidates, MTR incorporates spatial intention priors by adopting a small set of learnable motion query pairs. Each motion query pair takes charge of trajectory prediction and refinement for a specific motion mode, which stabilizes the training process and facilitates better multimodal predictions. Experiments show that MTR achieves state-of-the-art performance on both the marginal and joint motion prediction challenges, ranking 1st on the leaderbaords of Waymo Open Motion Dataset. Code will be available at https://github.com/sshaoshuai/MTR.

Semi-supervised learning (SSL) improves model generalization by leveraging massive unlabeled data to augment limited labeled samples. However, currently, popular SSL evaluation protocols are often constrained to computer vision (CV) tasks. In addition, previous work typically trains deep neural networks from scratch, which is time-consuming and environmentally unfriendly. To address the above issues, we construct a Unified SSL Benchmark (USB) for classification by selecting 15 diverse, challenging, and comprehensive tasks from CV, natural language processing (NLP), and audio processing (Audio), on which we systematically evaluate the dominant SSL methods, and also open-source a modular and extensible codebase for fair evaluation of these SSL methods. We further provide the pre-trained versions of the state-of-the-art neural models for CV tasks to make the cost affordable for further tuning. USB enables the evaluation of a single SSL algorithm on more tasks from multiple domains but with less cost. Specifically, on a single NVIDIA V100, only 39 GPU days are required to evaluate FixMatch on 15 tasks in USB while 335 GPU days (279 GPU days on 4 CV datasets except for ImageNet) are needed on 5 CV tasks with TorchSSL.

Since out-of-distribution generalization is a generally ill-posed problem, various proxy targets (e.g., calibration, adversarial robustness, algorithmic corruptions, invariance across shifts) were studied across different research programs resulting in different recommendations. While sharing the same aspirational goal, these approaches have never been tested under the same experimental conditions on real data. In this paper, we take a unified view of previous work, highlighting message discrepancies that we address empirically, and providing recommendations on how to measure the robustness of a model and how to improve it. To this end, we collect 172 publicly available dataset pairs for training and out-of-distribution evaluation of accuracy, calibration error, adversarial attacks, environment invariance, and synthetic corruptions. We fine-tune over 31k networks, from nine different architectures in the many- and few-shot setting. Our findings confirm that in- and out-of-distribution accuracies tend to increase jointly, but show that their relation is largely dataset-dependent, and in general more nuanced and more complex than posited by previous, smaller scale studies.

Abstract Despite their recent successes, generative adversarial networks (GANs) for semantic image synthesis still suffer from poor image quality when trained with only adversarial supervision. Previously, additionally employing the VGG-based perceptual loss has helped to overcome this issue, significantly improving the synthesis quality, but at the same time limited the progress of GAN models for semantic image synthesis. In this work, we propose a novel, simplified GAN model, which needs only adversarial supervision to achieve high quality results. We re-design the discriminator as a semantic segmentation network, directly using the given semantic label maps as the ground truth for training. By providing stronger supervision to the discriminator as well as to the generator through spatially- and semantically-aware discriminator feedback, we are able to synthesize images of higher fidelity and with a better alignment to their input label maps, making the use of the perceptual loss superfluous. Furthermore, we enable high-quality multi-modal image synthesis through global and local sampling of a 3D noise tensor injected into the generator, which allows complete or partial image editing. We show that images synthesized by our model are more diverse and follow the color and texture distributions of real images more closely. We achieve a strong improvement in image synthesis quality over prior state-of-the-art models across the commonly used ADE20K, Cityscapes, and COCO-Stuff datasets using only adversarial supervision. In addition, we investigate semantic image synthesis under severe class imbalance and sparse annotations, which are common aspects in practical applications but were overlooked in prior works. To this end, we evaluate our model on LVIS, a dataset originally introduced for long-tailed object recognition. We thereby demonstrate high performance of our model in the sparse and unbalanced data regimes, achieved by means of the proposed 3D noise and the ability of our discriminator to balance class contributions directly in the loss function. Our code and pretrained models are available at https://github.com/boschresearch/OASIS .

We present a simple and effective approach for non-blind image deblurring, combining classical techniques and deep learning. In contrast to existing methods that deblur the image directly in the standard image space, we propose to perform an explicit deconvolution process in a feature space by integrating a classical Wiener deconvolution framework with learned deep features. A multi-scale cascaded feature refinement module then predicts the deblurred image from the deconvolved deep features, progressively recovering detail and small-scale structures. The proposed model is trained in an end-to-end manner and evaluated on scenarios with simulated Gaussian noise, saturated pixels, or JPEG compression artifacts as well as real-world images. Moreover, we present detailed analyses of the benefit of the feature-based Wiener deconvolution and of the multi-scale cascaded feature refinement as well as the robustness of the proposed approach. Our extensive experimental results show that the proposed deep Wiener deconvolution network facilitates deblurred results with visibly fewer artifacts and quantitatively outperforms state-of-the-art non-blind image deblurring methods by a wide margin.

Meta-learning has been proposed as a framework to address the challenging few-shot learning setting. The key idea is to leverage a large number of similar few-shot tasks in order to learn how to adapt a base-learner to a new task for which only a few labeled samples are available. As deep neural networks (DNNs) tend to overfit using a few samples only, typical meta-learning models use shallow neural networks, thus limiting its effectiveness. In order to achieve top performance, some recent works tried to use the DNNs pre-trained on large-scale datasets but mostly in straight-forward manners, e.g., (1) taking their weights as a warm start of meta-training, and (2) freezing their convolutional layers as the feature extractor of base-learners. In this paper, we propose a novel approach called meta-transfer learning (MTL), which learns to transfer the weights of a deep NN for few-shot learning tasks. Specifically, meta refers to training multiple tasks, and transfer is achieved by learning scaling and shifting functions of DNN weights (and biases) for each task. To further boost the learning efficiency of MTL, we introduce the hard task (HT) meta-batch scheme as an effective learning curriculum of few-shot classification tasks. We conduct experiments for five-class few-shot classification tasks on three challenging benchmarks, miniImageNet, tieredImageNet, and Fewshot-CIFAR100 (FC100), in both supervised and semi-supervised settings. Extensive comparisons to related works validate that our MTL approach trained with the proposed HT meta-batch scheme achieves top performance. An ablation study also shows that both components contribute to fast convergence and high accuracy.

Few-shot learning aims to recognize novel classes from a few examples. Although significant progress has been made in the image domain, few-shot video classification is relatively unexplored. We argue that previous methods underestimate the importance of video feature learning and propose to learn spatiotemporal features using a 3D CNN. Proposing a two-stage approach that learns video features on base classes followed by fine-tuning the classifiers on novel classes, we show that this simple baseline approach outperforms prior few-shot video classification methods by over 20 points on existing benchmarks. To circumvent the need of labeled examples, we present two novel approaches that yield further improvement. First, we leverage tag-labeled videos from a large dataset using tag retrieval followed by selecting the best clips with visual similarities. Second, we learn generative adversarial networks that generate video features of novel classes from their semantic embeddings. Moreover, we find existing benchmarks are limited because they only focus on 5 novel classes in each testing episode and introduce more realistic benchmarks by involving more novel classes, i.e., few-shot learning, as well as a mixture of novel and base classes, i.e., generalized few-shot learning. The experimental results show that our retrieval and feature generation approach significantly outperform the baseline approach on the new benchmarks.

Integrated interpretability without sacrificing the prediction accuracy of decision making algorithms has the potential of greatly improving their value to the user. Instead of assigning a label to an image directly, we propose to learn iterative binary sub-decisions, inducing sparsity and transparency in the decision making process. The key aspect of our model is its ability to build a decision tree whose structure is encoded into the memory representation of a Recurrent Neural Network jointly learned by two models communicating through message passing. In addition, our model assigns a semantic meaning to each decision in the form of binary attributes, providing concise, semantic and relevant rationalizations to the user. On three benchmark image classification datasets, including the large-scale ImageNet, our model generates human interpretable binary decision sequences explaining the predictions of the network while maintaining state-of-the-art accuracy.

Accurate prediction of pedestrian and bicyclist paths is integral to the development of reliable autonomous vehicles in dense urban environments. The interactions between vehicle and pedestrian or bicyclist have a significant impact on the trajectories of traffic participants e.g. stopping or turning to avoid collisions. Although recent datasets and trajectory prediction approaches have fostered the development of autonomous vehicles yet the amount of vehicle-pedestrian (bicyclist) interactions modeled are sparse. In this work, we propose Euro-PVI, a dataset of pedestrian and bicyclist trajectories. In particular, our dataset caters more diverse and complex interactions in dense urban scenarios compared to the existing datasets. To address the challenges in predicting future trajectories with dense interactions, we develop a joint inference model that learns an expressive multi-modal shared latent space across agents in the urban scene. This enables our Joint-β-cVAE approach to better model the distribution of future trajectories. We achieve state of the art results on the nuScenes and Euro-PVI datasets demonstrating the importance of capturing interactions between ego-vehicle and pedestrians (bicyclists) for accurate predictions.

We introduce a new family of neural network models called Convolutional Dynamic Alignment Networks1(CoDA-Nets), which are performant classifiers with a high degree of inherent interpretability. Their core building blocks are Dynamic Alignment Units (DAUs), which linearly transform their input with weight vectors that dynamically align with task-relevant patterns. As a result, CoDA-Nets model the classification prediction through a series of input-dependent linear transformations, allowing for linear decomposition of the output into individual input contributions. Given the alignment of the DAUs, the resulting contribution maps align with discriminative input patterns. These model-inherent decompositions are of high visual quality and outperform existing attribution methods under quantitative metrics. Further, CoDA-Nets constitute performant classifiers, achieving on par results to ResNet and VGG models on e.g. CIFAR-10 and TinyImagenet.

The classical maximum a-posteriori (MAP) framework for non-blind image deblurring requires defining suitable data and regularization terms, whose interplay yields the desired clear image through optimization. The vast majority of prior work focuses on advancing one of these two crucial ingredients, while keeping the other one standard. Considering the indispensable roles and interplay of both data and regularization terms, we propose a simple and effective approach to jointly learn these two terms, embedding deep neural networks within the constraints of the MAP framework, trained in an end-to-end manner. The neural networks not only yield suitable image-adaptive features for both terms, but actually predict per-pixel spatially-variant features instead of the commonly used spatially-uniform ones. The resulting spatially-variant data and regularization terms particularly improve the restoration of fine-scale structures and detail. Quantitative and qualitative results underline the effectiveness of our approach, substantially outperforming the current state of the art.

Class-Incremental Learning (CIL) aims to learn a classification model with the number of classes increasing phase-by-phase. An inherent problem in CIL is the stability-plasticity dilemma between the learning of old and new classes, i.e., high-plasticity models easily forget old classes, but high-stability models are weak to learn new classes. We alleviate this issue by proposing a novel network architecture called Adaptive Aggregation Networks (AANets) in which we explicitly build two types of residual blocks at each residual level (taking ResNet as the baseline architecture): a stable block and a plastic block. We aggregate the output feature maps from these two blocks and then feed the results to the next-level blocks. We adapt the aggregation weights in order to balance these two types of blocks, i.e., to balance stability and plasticity, dynamically. We conduct extensive experiments on three CIL benchmarks: CIFAR-100, ImageNet-Subset, and ImageNet, and show that many existing CIL methods can be straightforwardly incorporated into the architecture of AANets to boost their performances1.

Both generalized and incremental few-shot learning have to deal with three major challenges: learning novel classes from only few samples per class, preventing catastrophic forgetting of base classes, and classifier calibration across novel and base classes. In this work we propose a three-stage framework that allows to explicitly and effectively address these challenges. While the first phase learns base classes with many samples, the second phase learns a calibrated classifier for novel classes from few samples while also preventing catastrophic forgetting. In the final phase, calibration is achieved across all classes. We evaluate the proposed framework on four challenging benchmark datasets for image and video few-shot classification and obtain state-of-the-art results for both generalized and incremental few shot learning.

In order to robustly deploy object detectors across a wide range of scenarios, they should be adaptable to shifts in the input distribution without the need to constantly an-notate new data. This has motivated research in Unsupervised Domain Adaptation (UDA) algorithms for detection. UDA methods learn to adapt from labeled source domains to unlabeled target domains, by inducing alignment between detector features from source and target domains. Yet, there is no consensus on what features to align and how to do the alignment. In our work, we propose a framework that generalizes the different components commonly used by UDA methods laying the ground for an in-depth analysis of the UDA design space. Specifically, we propose a novel UDA algorithm, ViSGA, a direct implementation of our framework, that leverages the best design choices and introduces a simple but effective method to aggregate features at instance-level based on visual similarity before inducing group alignment via adversarial training. We show that both similarity-based grouping and adversarial training allows our model to focus on coarsely aligning feature groups, without being forced to match all instances across loosely aligned domains. Finally, we examine the applicability of ViSGA to the setting where labeled data are gathered from different sources. Experiments show that not only our method outperforms previous single-source approaches on Sim2Real and Adverse Weather, but also generalizes well to the multi-source setting.

Adversarial training (AT) has become the de-facto standard to obtain models robust against adversarial examples. However, AT exhibits severe robust overfitting: cross-entropy loss on adversarial examples, so-called robust loss, decreases continuously on training examples, while eventually increasing on test examples. In practice, this leads to poor robust generalization, i.e., adversarial robustness does not generalize well to new examples. In this paper, we study the relationship between robust generalization and flatness of the robust loss landscape in weight space, i.e., whether robust loss changes significantly when perturbing weights. To this end, we propose average- and worst-case metrics to measure flatness in the robust loss landscape and show a correlation between good robust generalization and flatness. For example, throughout training, flatness reduces significantly during overfitting such that early stopping effectively finds flatter minima in the robust loss landscape. Similarly, AT variants achieving higher adversarial robustness also correspond to flatter minima. This holds for many popular choices, e.g., AT-AWP, TRADES, MART, AT with self-supervision or additional unlabeled examples, as well as simple regularization techniques, e.g., AutoAugment, weight decay or label noise. For fair comparison across these approaches, our flatness measures are specifically designed to be scale-invariant and we conduct extensive experiments to validate our findings.

Class-Incremental Learning (CIL) [38] trains classifiers under a strict memory budget: in each incremental phase, learning is done for new data, most of which is abandoned to free space for the next phase. The preserved data are exemplars used for replaying. However, existing methods use a static and ad hoc strategy for memory allocation, which is often sub-optimal. In this work, we propose a dynamic memory management strategy that is optimized for the incremental phases and different object classes. We call our method reinforced memory management (RMM), leveraging reinforcement learning. RMM training is not naturally compatible with CIL as the past, and future data are strictly non-accessible during the incremental phases. We solve this by training the policy function of RMM on pseudo CIL tasks, e.g., the tasks built on the data of the zeroth phase, and then applying it to target tasks. RMM propagates two levels of actions: Level-1 determines how to split the memory between old and new classes, and Level-2 allocates memory for each specific class. In essence, it is an optimizable and general method for memory management that can be used in any replaying-based CIL method. For evaluation, we plug RMM into two top-performing baselines (LUCIR+AANets and POD+AANets [28]) and conduct experiments on three benchmarks (CIFAR-100, ImageNet-Subset, and ImageNet-Full). Our results show clear improvements, e.g., boosting POD+AANets by 3.6%, 4.4%, and 1.9% in the 25-Phase settings of the above benchmarks, respectively. The code is available at https://class-il.mpi-inf.mpg.de/rmm/.

In this paper, we aim to tackle the task of Assessing Future Moment of an Action of Interest (AFM-AI). The goal of this task is to assess if an action of interest will happen or not as well as the starting moment of the action. We aim to assess starting moments at any time-horizon of the future. To this end, we tackle the regression task of the starting moments as a generation task using a Deterministic Residual Guided Variational Regression Module (DR-VRM), which is built on a Variational Regression Module (VRM) and a deterministic residual network. The VRM takes the uncertainty into account and is capable of generating diverse predictions for the starting moment. The deterministic network encourages the VRM to learn from deterministic residual information in order to generate more precise predictions for moment assessment. Experimental results on three datasets clearly show that the proposed method is capable of generating both diverse and precise predictions of starting moments for query actions.

Abstract Today’s deep learning systems deliver high performance based on end-to-end training but are notoriously hard to inspect. We argue that there are at least two reasons making inspectability challenging: (i) representations are distributed across hundreds of channels and (ii) a unifying metric quantifying inspectability is lacking. In this paper, we address both issues by proposing Semantic Bottlenecks (SB), which can be integrated into pretrained networks, to align channel outputs with individual visual concepts and introduce the model agnostic Area Under inspectability Curve (AUiC) metric to measure the alignment. We present a case study on semantic segmentation to demonstrate that SBs improve the AUiC up to six-fold over regular network outputs. We explore two types of SB-layers in this work. First, concept-supervised SB-layers (SSB), which offer inspectability w.r.t. predefined concepts that the model is demanded to rely on. And second, unsupervised SBs (USB), which offer equally strong AUiC improvements by restricting distributedness of representations across channels. Importantly, for both SB types, we can recover state of the art segmentation performance across two different models despite a drastic dimensionality reduction from 1000s of non aligned channels to 10s of semantics-aligned channels that all downstream results are based on.

Person Search is a challenging task which requires to retrieve a person's image and the corresponding position from an image dataset. It consists of two sub-tasks: pedestrian detection and person re-identification (re-ID). One of the key challenges is to properly combine the two sub-tasks into a unified framework. Existing works usually adopt a straightforward strategy by concatenating a detector and a re-ID model directly, either into an integrated model or into separated models. We argue that simply concatenating detection and re-ID is a sub-optimal solution, and we propose a Hierarchical Online Instance Matching (HOIM) loss which exploits the hierarchical relationship between detection and re-ID to guide the learning of our network. Our novel HOIM loss function harmonizes the objectives of the two sub-tasks and encourages better feature learning. In addition, we improve the loss update policy by introducing Selective Memory Refreshment (SMR) for unlabeled persons, which takes advantage of the potential discrimination power of unlabeled data. From the experiments on two standard person search benchmarks, i.e. CUHK-SYSU and PRW, we achieve state-of-the-art performance, which justifies the effectiveness of our proposed HOIM loss on learning robust features.

A problem in automatically generated stories for image sequences is that they use overly generic vocabulary and phrase structure and fail to match the distributional characteristics of human-generated text. We address this problem by introducing explicit representations for objects and their relations by extracting scene graphs from the images. Utilizing an embedding of this scene graph enables our model to more explicitly reason over objects and their relations during story generation, compared to the global features from an object classifier used in previous work. We apply metrics that account for the diversity of words and phrases of generated stories as well as for reference to narratively-salient image features and show that our approach outperforms previous systems. Our experiments also indicate that our models obtain competitive results on reference-based metrics.

Flow-based generative models are an important class of exact inference models that admit efficient inference and sampling for image synthesis. Owing to the efficiency constraints on the design of the flow layers, e.g. split coupling flow layers in which approximately half the pixels do not undergo further transformations, they have limited expressiveness for modeling long-range data dependencies compared to autoregressive models that rely on conditional pixel-wise generation. In this work, we improve the representational power of flow-based models by introducing channel-wise dependencies in their latent space through multi-scale autoregressive priors (mAR). Our mAR prior for models with split coupling flow layers (mAR-SCF) can better capture dependencies in complex multimodal data. The resulting model achieves state-of-the-art density estimation results on MNIST, CIFAR-10, and ImageNet. Furthermore, we show that mAR-SCF allows for improved image generation quality, with gains in FID and Inception scores compared to state-of-the-art flow-based models.

Person Search is a practically relevant task that aims to jointly solve Person Detection and Person Re-identification (re-ID). Specifically, it requires to find and locate all instances with the same identity as the query person in a set of panoramic gallery images. One major challenge comes from the contradictory goals of the two sub-tasks, i.e., person detection focuses on finding the commonness of all persons while person re-ID handles the differences among multiple identities. Therefore, it is crucial to reconcile the relationship between the two sub-tasks in a joint person search model. To this end, We present a novel approach called Norm-Aware Embedding to disentangle the person embedding into norm and angle for detection and re-ID respectively, allowing for both effective and efficient multi-task training. We further extend the proposal-level person embedding to pixel-level, whose discrimination ability is less affected by mis-alignment. We outperform other one-step methods by a large margin and achieve comparable performance to two-step methods on both CUHK-SYSU and PRW. Also, Our method is easy to train and resource-friendly, running at 12 fps on a single GPU.

Multi-Class Incremental Learning (MCIL) aims to learn new concepts by incrementally updating a model trained on previous concepts. However, there is an inherent trade-off to effectively learning new concepts without catastrophic forgetting of previous ones. To alleviate this issue, it has been proposed to keep around a few examples of the previous concepts but the effectiveness of this approach heavily depends on the representativeness of these examples. This paper proposes a novel and automatic framework we call mnemonics, where we parameterize exemplars and make them optimizable in an end-to-end manner. We train the framework through bilevel optimizations, i.e., model-level and exemplar-level. We conduct extensive experiments on three MCIL benchmarks, CIFAR-100, ImageNet-Subset and ImageNet, and show that using mnemonics exemplars can surpass the state-of-the-art by a large margin. Interestingly and quite intriguingly, the mnemonics exemplars tend to be on the boundaries between different classes.

Among the major remaining challenges for generative adversarial networks (GANs) is the capacity to synthesize globally and locally coherent images with object shapes and textures indistinguishable from real images. To target this issue we propose an alternative U-Net based discriminator architecture, borrowing the insights from the segmentation literature. The proposed U-Net based architecture allows to provide detailed per-pixel feedback to the generator while maintaining the global coherence of synthesized images, by providing the global image feedback as well. Empowered by the per-pixel response of the discriminator, we further propose a per-pixel consistency regularization technique based on the CutMix data augmentation, encouraging the U-Net discriminator to focus more on semantic and structural changes between real and fake images. This improves the U-Net discriminator training, further enhancing the quality of generated samples. The novel discriminator improves over the state of the art in terms of the standard distribution and image quality metrics, enabling the generator to synthesize images with varying structure, appearance and levels of detail, maintaining global and local realism. Compared to the BigGAN baseline, we achieve an average improvement of 2.7 FID points across FFHQ, CelebA, and the proposed COCO-Animals dataset.

Adversarial training yields robust models against a specific threat model, e.g., $L_\infty$ adversarial examples. Typically robustness does not generalize to previously unseen threat models, e.g., other $L_p$ norms, or larger perturbations. Our confidence-calibrated adversarial training (CCAT) tackles this problem by biasing the model towards low confidence predictions on adversarial examples. By allowing to reject examples with low confidence, robustness generalizes beyond the threat model employed during training. CCAT, trained only on $L_\infty$ adversarial examples, increases robustness against larger $L_\infty$, $L_2$, $L_1$ and $L_0$ attacks, adversarial frames, distal adversarial examples and corrupted examples and yields better clean accuracy compared to adversarial training. For thorough evaluation we developed novel white- and black-box attacks directly attacking CCAT by maximizing confidence. For each threat model, we use $7$ attacks with up to $50$ restarts and $5000$ iterations and report worst-case robust test error, extended to our confidence-thresholded setting, across all attacks.

We present a simple and effective approach for non-blind image deblurring, combining classical techniques and deep learning. In contrast to existing methods that deblur the image directly in the standard image space, we propose to perform an explicit deconvolution process in a feature space by integrating a classical Wiener deconvolution framework with learned deep features. A multi-scale feature refinement module then predicts the deblurred image from the deconvolved deep features, progressively recovering detail and small-scale structures. The proposed model is trained in an end-to-end manner and evaluated on scenarios with both simulated and real-world image blur. Our extensive experimental results show that the proposed deep Wiener deconvolution network facilitates deblurred results with visibly fewer artifacts. Moreover, our approach quantitatively outperforms state-of-the-art non-blind image deblurring methods by a wide margin.

From the beginning of zero-shot learning research, visual attributes have been shown to play an important role. In order to better transfer attribute-based knowledge from known to unknown classes, we argue that an image representation with integrated attribute localization ability would be beneficial for zero-shot learning. To this end, we propose a novel zero-shot representation learning framework that jointly learns discriminative global and local features using only class-level attributes. While a visual-semantic embedding layer learns global features, local features are learned through an attribute prototype network that simultaneously regresses and decorrelates attributes from intermediate features. We show that our locality augmented image representations achieve a new state-of-the-art on three zero-shot learning benchmarks. As an additional benefit, our model points to the visual evidence of the attributes in an image, e.g. for the CUB dataset, confirming the improved attribute localization ability of our image representation.

Models for computer vision are commonly defined either w.r.t. low-level concepts such as pixels that are to be grouped, or w.r.t. high-level concepts such as semantic objects that are to be detected and tracked. Combining bottom-up grouping with top-down detection and tracking, although highly desirable, is a challenging problem. We state this joint problem as a co-clustering problem that is principled and tractable by existing algorithms. We demonstrate the effectiveness of this approach by combining bottom-up motion segmentation by grouping of point trajectories with high-level multiple object tracking by clustering of bounding boxes. We show that solving the joint problem is beneficial at the low-level, in terms of the FBMS59 motion segmentation benchmark, and at the high-level, in terms of the Multiple Object Tracking benchmarks MOT15, MOT16 and the MOT17 challenge, and is state-of-the-art in some metrics.

People nowadays share large parts of their personal lives through social media. Being able to automatically recognise people in personal photos may greatly enhance user convenience by easing photo album organisation. For human identification task, however, traditional focus of computer vision has been face recognition and pedestrian re-identification. Person recognition in social media photos sets new challenges for computer vision, including non-cooperative subjects (e.g. backward viewpoints, unusual poses) and great changes in appearance. To tackle this problem, we build a simple person recognition framework that leverages convnet features from multiple image regions (head, body, etc.). We propose new recognition scenarios that focus on the time and appearance gap between training and testing samples. We present an in-depth analysis of the importance of different features according to time and viewpoint generalisability. In the process, we verify that our simple approach achieves the state of the art result on the PIPA [1] benchmark, arguably the largest social media based benchmark for person recognition to date with diverse poses, viewpoints, social groups, and events. Compared the conference version of the paper [2], this paper additionally presents (1) analysis of a face recogniser (DeepID2+ [3]), (2) new method naeil2 that combines the conference version method naeil and DeepID2+ to achieve state of the art results even compared to post-conference works, (3) discussion of related work since the conference version, (4) additional analysis including the head viewpoint-wise breakdown of performance, and (5) results on the open-world setup.

The goal of human action anticipation is to predict future actions. Ideally, in real-world applications such as video surveillance and self-driving systems, future actions should not only be predicted with high accuracy but also at arbitrary and variable time-horizons ranging from short- to long-term predictions. Current work mostly focuses on predicting the next action and thus long-term prediction is achieved by recursive prediction of each next action, which is both inefficient and accumulates errors. In this paper, we propose a novel time-conditioned method for efficient and effective long-term action anticipation. There are two key ingredients to our approach. First, by explicitly conditioning our anticipation network on time allows to efficiently anticipate also long-term actions. And second, we propose an attended temporal feature and a time-conditioned skip connection to extract relevant and useful information from observations for effective anticipation. We conduct extensive experiments on the large-scale Epic-Kitchen and the 50Salads Datasets. Experimental results show that the proposed method is capable of anticipating future actions at both short-term and long-term, and achieves state-of-the-art performance.

Machine Learning (ML) models are increasingly deployed in the wild to perform a wide range of tasks. In this work, we ask to what extent can an adversary steal functionality of such ``victim'' models based solely on blackbox interactions: image in, predictions out. In contrast to prior work, we study complex victim blackbox models, and an adversary lacking knowledge of train/test data used by the model, its internals, and semantics over model outputs. We formulate model functionality stealing as a two-step approach: (i) querying a set of input images to the blackbox model to obtain predictions; and (ii) training a ``knockoff'' with queried image-prediction pairs. We make multiple remarkable observations: (a) querying random images from a different distribution than that of the blackbox training data results in a well-performing knockoff; (b) this is possible even when the knockoff is represented using a different architecture; and (c) our reinforcement learning approach additionally improves query sample efficiency in certain settings and provides performance gains. We validate model functionality stealing on a range of datasets and tasks, as well as show that a reasonable knockoff of an image analysis API could be created for as little as $30.

Importance of visual context in scene understanding tasks is well recognized in the computer vision community. However, to what extent the computer vision models are dependent on the context to make their predictions is unclear. A model overly relying on context will fail when encountering objects in different contexts than in training data and hence it is important to identify these dependencies before we can deploy the models in the real-world. We propose a method to quantify the sensitivity of black-box vision models to visual context by editing images to remove selected objects and measuring the response of the target models. We apply this methodology on two tasks, image classification and semantic segmentation, and discover undesirable dependency between objects and context, for example that ``sidewalk'' segmentation is very sensitive to the presence of ``cars'' in the image. We propose an object removal based data augmentation solution to mitigate this dependency and increase the robustness of classification and segmentation models to contextual variations. Our experiments show that the proposed data augmentation helps these models improve the performance in out-of-context scenarios, while preserving the performance on regular data.

Obtaining deep networks that are robust against adversarial examples and generalize well is an open problem. A recent hypothesis even states that both robust and accurate models are impossible, i.e., adversarial robustness and generalization are conflicting goals. In an effort to clarify the relationship between robustness and generalization, we assume an underlying, low-dimensional data manifold and show that: 1. regular adversarial examples leave the manifold; 2. adversarial examples constrained to the manifold, i.e., on-manifold adversarial examples, exist; 3. on-manifold adversarial examples are generalization errors, and on-manifold adversarial training boosts generalization; 4. regular robustness and generalization are not necessarily contradicting goals. These assumptions imply that both robust and accurate models are possible. However, different models (architectures, training strategies etc.) can exhibit different robustness and generalization characteristics. To confirm our claims, we present extensive experiments on synthetic data (with known manifold) as well as on EMNIST, Fashion-MNIST and CelebA.

Meta-learning has been proposed as a framework to address the challenging few-shot learning setting. The key idea is to leverage a large number of similar few-shot tasks in order to learn how to adapt a base-learner to a new task for which only a few labeled samples are available. As deep neural networks (DNNs) tend to overfit using a few samples only, meta-learning typically uses shallow neural networks (SNNs), thus limiting its effectiveness. In this paper we propose a novel few-shot learning method called meta-transfer learning (MTL) which learns to adapt a deep NN for few shot learning tasks. Specifically, "meta" refers to training multiple tasks, and "transfer" is achieved by learning scaling and shifting functions of DNN weights for each task. In addition, we introduce the hard task (HT) meta-batch scheme as an effective learning curriculum for MTL. We conduct experiments using (5-class, 1-shot) and (5-class, 5-shot) recognition tasks on two challenging few-shot learning benchmarks: miniImageNet and Fewshot-CIFAR100. Extensive comparisons to related works validate that our meta-transfer learning approach trained with the proposed HT meta-batch scheme achieves top performance. An ablation study also shows that both components contribute to fast convergence and high accuracy.

Semantic segmentation is one of the most fundamental problems in computer vision and pixel-level labelling in this context is particularly expensive. Hence, there have been several attempts to reduce the annotation effort such as learning from image level labels and bounding box annotations. In this paper we take this one step further and focus on the challenging task of zero- and few-shot learning of semantic segmentation. We define this task as image segmentation by assigning a label to every pixel even though either no labeled sample of that class was present during training, i.e. zero-label semantic segmentation, or only a few labeled samples were present, i.e. few-label semantic segmentation.Our goal is to transfer the knowledge from previously seen classes to novel classes. Our proposed semantic projection network (SPNet) achieves this goal by incorporating a class-level semantic information into any network designed for semantic segmentation, in an end-to-end manner. We also propose a benchmark for this task on the challenging COCO-Stuff and PASCAL VOC12 datasets. Our model is effective in segmenting novel classes, i.e. alleviating expensive dense annotations, but also in adapting to novel classes without forgetting its prior knowledge, i.e. generalized zero- and few-label semantic segmentation.

When labeled training data is scarce, a promising data augmentation approach is to generate visual features of unknown classes using their attributes. To learn the class conditional distribution of CNN features, these models rely on pairs of image features and class attributes. Hence, they can not make use of the abundance of unlabeled data samples. In this paper, we tackle any-shot learning problems i.e. zero-shot and few-shot, in a unified feature generating framework that operates in both inductive and transductive learning settings. We develop a conditional generative model that combines the strength of VAE and GANs and in addition, via an unconditional discriminator, learns the marginal feature distribution of unlabeled images. We empirically show that our model learns highly discriminative CNN features for five datasets, i.e. CUB, SUN, AWA and ImageNet, and establish a new state-of-the-art in any-shot learning, i.e. inductive and transductive (generalized) zero- and few-shot learning settings. We also demonstrate that our learned features are interpretable: we visualize them by inverting them back to the pixel space and we explain them by generating textual arguments of why they are associated with a certain label.

Few-shot classification (FSC) is challenging due to the scarcity of labeled training data (e.g. only one labeled data point per class). Meta-learning has shown to achieve promising results by learning to initialize a classification model for FSC. In this paper we propose a novel semi-supervised meta-learning method called learning to self-train (LST) that leverages unlabeled data and specifically meta-learns how to cherry-pick and label such unsupervised data to further improve performance. To this end, we train the LST model through a large number of semi-supervised few-shot tasks. On each task, we train a few-shot model to predict pseudo labels for unlabeled data, and then iterate the self-training steps on labeled and pseudo-labeled data with each step followed by fine-tuning. We additionally learn a soft weighting network (SWN) to optimize the self-training weights of pseudo labels so that better ones can contribute more to gradient descent optimization. We evaluate our LST method on two ImageNet benchmarks for semi-supervised few-shot classification and achieve large improvements over the state-of-the-art.

Convolutional networks reach top quality in pixel-level video object segmentation but require a large amount of training data (1k–100k) to deliver such results. We propose a new training strategy which achieves state-of-the-art results across three evaluation datasets while using \(20\,\times \) – \(1000\,\times \) less annotated data than competing methods. Our approach is suitable for both single and multiple object segmentation. Instead of using large training sets hoping to generalize across domains, we generate in-domain training data using the provided annotation on the first frame of each video to synthesize—“lucid dream” (in a lucid dream the sleeper is aware that he or she is dreaming and is sometimes able to control the course of the dream)—plausible future video frames. In-domain per-video training data allows us to train high quality appearance- and motion-based models, as well as tune the post-processing stage. This approach allows to reach competitive results even when training from only a single annotated frame, without ImageNet pre-training. Our results indicate that using a larger training set is not automatically better, and that for the video object segmentation task a smaller training set that is closer to the target domain is more effective. This changes the mindset regarding how many training samples and general “objectness” knowledge are required for the video object segmentation task.

Due to the importance of zero-shot learning, i.e., classifying images where there is a lack of labeled training data, the number of proposed approaches has recently increased steadily. We argue that it is time to take a step back and to analyze the status quo of the area. The purpose of this paper is three-fold. First, given the fact that there is no agreed upon zero-shot learning benchmark, we first define a new benchmark by unifying both the evaluation protocols and data splits of publicly available datasets used for this task. This is an important contribution as published results are often not comparable and sometimes even flawed due to, e.g., pre-training on zero-shot test classes. Moreover, we propose a new zero-shot learning dataset, the Animals with Attributes 2 (AWA2) dataset which we make publicly available both in terms of image features and the images themselves. Second, we compare and analyze a significant number of the state-of-the-art methods in depth, both in the classic zero-shot setting but also in the more realistic generalized zero-shot setting. Finally, we discuss in detail the limitations of the current status of the area which can be taken as a basis for advancing it.

Direct prediction of 3D body pose and shape remains a challenge even for highly parameterized deep learning models. Mapping from the 2D image space to the prediction space is difficult: perspective ambiguities make the loss function noisy and training data is scarce. In this paper, we propose a novel approach (Neural Body Fitting (NBF)). It integrates a statistical body model within a CNN, leveraging reliable bottom-up semantic body part segmentation and robust top-down body model constraints. NBF is fully differentiable and can be trained using 2D and 3D annotations. In detailed experiments, we analyze how the components of our model affect performance, especially the use ofpart segmentations as an explicit intermediate representation, and present a robust, efficiently trainable framework for 3D human pose estimation from 2D images with competitive results on standard benchmarks. Code will be made available at http://github.com/mohomran/ neural_body_fitting.

Boundary estimation in images and videos has been a very active topic of research, and organizing visual information into boundaries and segments is believed to be a corner stone of visual perception. While prior work has focused on estimating boundaries for observed frames, our work aims at predicting boundaries of future unobserved frames. This requires our model to learn about the fate of boundaries and corresponding motion patterns---including a notion of "intuitive physics." We experiment on natural video sequences along with synthetic sequences with deterministic physics-based and agent-based motions. While not being our primary goal, we also show that fusion of RGB and boundary prediction leads to improved RGB predictions.

Existing systems for video-based pose estimation and tracking struggle to perform well on realistic videos with multiple people and often fail to output body-pose trajectories consistent over time. To address this shortcoming this paper introduces PoseTrack which is a new large-scale benchmark for video-based human pose estimation and articulated tracking. Our new benchmark encompasses three tasks focusing on i) single-frame multi-person pose estimation, ii) multi-person pose estimation in videos, and iii) multi-person articulated tracking. To establish the benchmark, we collect, annotate and release a new dataset that features videos with multiple people labeled with person tracks and articulated pose. A public centralized evaluation server is provided to allow the research community to evaluate on a held-out test set. Furthermore, we conduct an extensive experimental study on recent approaches to articulated pose tracking and provide analysis of the strengths and weaknesses of the state of the art. We envision that the proposed benchmark will stimulate productive research both by providing a large and representative training dataset as well as providing a platform to objectively evaluate and compare the proposed methods. The benchmark is freely accessible at https://posetrack.net/.

Progress towards advanced systems for assisted and autonomous driving is leveraging recent advances in recognition and segmentation methods. Yet, we are still facing challenges in bringing reliable driving to inner cities, as those are composed of highly dynamic scenes observed from a moving platform at considerable speeds. Anticipation becomes a key element in order to react timely and prevent accidents. In this paper we argue that it is necessary to predict at least 1 second and we thus propose a new model that jointly predicts ego motion and people trajectories over such large time horizons. We pay particular attention to modeling the uncertainty of our estimates arising from the non-deterministic nature of natural traffic scenes. Our experimental results show that it is indeed possible to predict people trajectories at the desired time horizons and that our uncertainty estimates are informative of the prediction error. We also show that both sequence modeling of trajectories as well as our novel method of long term odometry prediction are essential for best performance.

For autonomous agents to successfully operate in the real world, anticipation of future events and states of their environment is a key competence. This problem has been formalized as a sequence extrapolation problem, where a number of observations are used to predict the sequence into the future. Real-world scenarios demand a model of uncertainty of such predictions, as predictions become increasingly uncertain - in particular on long time horizons. While impressive results have been shown on point estimates, scenarios that induce multi-modal distributions over future sequences remain challenging. Our work addresses these challenges in a Gaussian Latent Variable model for sequence prediction. Our core contribution is a "Best of Many" sample objective that leads to more accurate and more diverse predictions that better capture the true variations in real-world sequence data. Beyond our analysis of improved model fit, our models also empirically outperform prior work on three diverse tasks ranging from traffic scenes to weather data.

Generating novel, yet realistic, images of persons is a challenging task due to the complex interplay between the different image factors, such as the foreground, background and pose information. In this work, we aim at generating such images based on a novel, two-stage reconstruction pipeline that learns a disentangled representation of the aforementioned image factors and generates novel person images at the same time. First, a multi-branched reconstruction network is proposed to disentangle and encode the three factors into embedding features, which are then combined to re-compose the input image itself. Second, three corresponding mapping functions are learned in an adversarial manner in order to map Gaussian noise to the learned embedding feature space, for each factor, respectively. Using the proposed framework, we can manipulate the foreground, background and pose of the input image, and also sample new embedding features to generate such targeted manipulations, that provide more control over the generation process. Experiments on the Market-1501 and Deepfashion datasets show that our model does not only generate realistic person images with new foregrounds, backgrounds and poses, but also manipulates the generated factors and interpolates the in-between states. Another set of experiments on Market-1501 shows that our model can also be beneficial for the person re-identification task1.

Images convey a broad spectrum of personal information. If such images are shared on social media platforms, this personal information is leaked which conflicts with the privacy of depicted persons. Therefore, we aim for automated approaches to redact such private information and thereby protect privacy of the individual. By conducting a user study we find that obfuscating the image regions related to the private information leads to privacy while retaining utility of the images. Moreover, by varying the size of the regions different privacy-utility trade-offs can be achieved. Our findings argue for a "redaction by segmentation" paradigm. Hence, we propose the first sizable dataset of private images "in the wild" annotated with pixel and instance level labels across a broad range of privacy classes. We present the first model for automatic redaction of diverse private information. It is effective at achieving various privacy-utility trade-offs within 83% of the performance of redactions based on ground-truth annotation.

Deep models that are both effective and explainable are desirable in many settings; prior explainable models have been unimodal, offering either image-based visualization of attention weights or text-based generation of post-hoc justifications. We propose a multimodal approach to explanation, and argue that the two modalities provide complementary explanatory strengths. We collect two new datasets to define and evaluate this task, and propose a novel model which can provide joint textual rationale generation and attention visualization. Our datasets define visual and textual justifications of a classification decision for activity recognition tasks (ACT-X) and for visual question answering tasks (VQA-X). We quantitatively show that training with the textual explanations not only yields better textual justification models, but also better localizes the evidence that supports the decision. We also qualitatively show cases where visual explanation is more insightful than textual explanation, and vice versa, supporting our thesis that multimodal explanation models offer significant benefits over unimodal approaches.

As more and more personal photos are shared online, being able to obfuscate identities in such photos is becoming a necessity for privacy protection. People have largely resorted to blacking out or blurring head regions, but they result in poor user experience while being surprisingly ineffective against state of the art person recognizers [17]. In this work, we propose a novel head inpainting obfuscation technique. Generating a realistic head inpainting in social media photos is challenging because subjects appear in diverse activities and head orientations. We thus split the task into two sub-tasks: (1) facial landmark generation from image context (e.g. body pose) for seamless hypothesis of sensible head pose, and (2) facial landmark conditioned head inpainting. We verify that our inpainting method generates realistic person images, while achieving superior obfuscation performance against automatic person recognizers.

Suffering from the extreme training data imbalance between seen and unseen classes, most of existing state-of-the-art approaches fail to achieve satisfactory results for the challenging generalized zero-shot learning task. To circumvent the need for labeled examples of unseen classes, we propose a novel generative adversarial network (GAN) that synthesizes CNN features conditioned on class-level semantic information, offering a shortcut directly from a semantic descriptor of a class to a class-conditional feature distribution. Our proposed approach, pairing a Wasserstein GAN with a classification loss, is able to generate sufficiently discriminative CNN features to train softmax classifiers or any multimodal embedding method. Our experimental results demonstrate a significant boost in accuracy over the state of the art on five challenging datasets - CUB, FLO, SUN, AWA and ImageNet - in both the zero-shot learning and generalized zero-shot learning settings.

Pedestrian detection has progressed significantly in the last years. However, occluded people are notoriously hard to detect, as their appearance varies substantially depending on a wide range of occlusion patterns. In this paper, we aim to propose a simple and compact method based on the FasterRCNN architecture for occluded pedestrian detection. We start with interpreting CNN channel features of a pedestrian detector, and we find that different channels activate responses for different body parts respectively. These findings motivate us to employ an attention mechanism across channels to represent various occlusion patterns in one single model, as each occlusion pattern can be formulated as some specific combination of body parts. Therefore, an attention network with self or external guidances is proposed as an add-on to the baseline FasterRCNN detector. When evaluating on the heavy occlusion subset, we achieve a significant improvement of 8pp to the baseline FasterRCNN detector on CityPersons and on Caltech we outperform the state-of-the-art method by 4pp.

While great progress has been made recently in automatic image manipulation, it has been limited to object centric images like faces or structured scene datasets. In this work, we take a step towards general scene-level image editing by developing an automatic interaction-free object removal model. Our model learns to find and remove objects from general scene images using image-level labels and unpaired data in a generative adversarial network (GAN) framework. We achieve this with two key contributions: a two-stage editor architecture consisting of a mask generator and image in-painter that co-operate to remove objects, and a novel GAN based prior for the mask generator that allows us to flexibly incorporate knowledge about object shapes. We experimentally show on two datasets that our method effectively removes a wide variety of objects using weak supervision only.

Top-k error is currently a popular performance measure on large scale image classification benchmarks such as ImageNet and Places. Despite its wide acceptance, our understanding of this metric is limited as most of the previous research is focused on its special case, the top-1 error. In this work, we explore two directions that shed more light on the top-k error. First, we provide an in-depth analysis of established and recently proposed single-label multiclass methods along with a detailed account of efficient optimization algorithms for them. Our results indicate that the softmax loss and the smooth multiclass SVM are surprisingly competitive in top-k error uniformly across all k, which can be explained by our analysis of multiclass top-k calibration. Further improvements for a specific k are possible with a number of proposed top-k loss functions. Second, we use the top-k methods to explore the transition from multiclass to multilabel learning. In particular, we find that it is possible to obtain effective multilabel classifiers on Pascal VOC using a single label per image for training, while the gap between multiclass and multilabel methods on MS COCO is more significant. Finally, our contribution of efficient algorithms for training with the considered top-k and multilabel loss functions is of independent interest.

Encouraged by the recent progress in pedestrian detection, we investigate the gap between current state-of-the-art methods and the "perfect single frame detector". We enable our analysis by creating a human baseline for pedestrian detection (over the Caltech pedestrian dataset). After manually clustering the frequent errors of a top detector, we characterise both localisation and background-versus-foreground errors. To address localisation errors we study the impact of training annotation noise on the detector performance, and show that we can improve results even with a small portion of sanitised training data. To address background/foreground discrimination, we study convnets for pedestrian detection, and discuss which factors affect their performance. Other than our in-depth analysis, we report top performance on the Caltech pedestrian dataset, and provide a new sanitised set of training and test annotations.

In this paper we propose an approach for articulated tracking of multiple people in unconstrained videos. Our starting point is a model that resembles existing architectures for single-frame pose estimation but is substantially faster. We achieve this in two ways: (1) by simplifying and sparsifying the body-part relationship graph and leveraging recent methods for faster inference, and (2) by offloading a substantial share of computation onto a feed-forward convolutional architecture that is able to detect and associate body joints of the same person even in clutter. We use this model to generate proposals for body joint locations and formulate articulated tracking as spatio-temporal grouping of such proposals. This allows to jointly solve the association problem for all people in the scene by propagating evidence from strong detections through time and enforcing constraints that each proposal can be assigned to one person only. We report results on a public MPII Human Pose benchmark and on a new MPII Video Pose dataset of image sequences with multiple people. We demonstrate that our model achieves state-of-the-art results while using only a fraction of time and is able to leverage temporal information to improve state-of-the-art for crowded scenes.

Zero-shot image classification using auxiliary information, such as attributes describing discriminative object properties, requires time-consuming annotation by domain experts. We instead propose a method that relies on human gaze as auxiliary information, exploiting that even non-expert users have a natural ability to judge class membership. We present a data collection paradigm that involves a discrimination task to increase the information content obtained from gaze data. Our method extracts discriminative descriptors from the data and learns a compatibility function between image and gaze using three novel gaze embeddings: Gaze Histograms (GH), Gaze Features with Grid (GFG) and Gaze Features with Sequence (GFS). We introduce two new gaze-annotated datasets for fine-grained image classification and show that human gaze data is indeed class discriminative, provides a competitive alternative to expert-annotated attributes, and outperforms other baselines for zero-shot image classification.

Semantic labelling and instance segmentation are two tasks that require particularly costly annotations. Starting from weak supervision in the form of bounding box detection annotations, we propose a new approach that does not require modification of the segmentation training procedure. We show that when carefully designing the input labels from given bounding boxes, even a single round of training is enough to improve over previously reported weakly supervised results. Overall, our weak supervision approach reaches ~95% of the quality of the fully supervised model, both for semantic labelling and instance segmentation.

We state a combinatorial optimization problem whose feasible solutions define both a decomposition and a node labeling of a given graph. This problem offers a common mathematical abstraction of seemingly unrelated computer vision tasks, including instance-separating semantic segmentation, articulated human body pose estimation and multiple object tracking. Conceptually, it generalizes the unconstrained integer quadratic program and the minimum cost lifted multicut problem, both of which are NP-hard. In order to find feasible solutions efficiently, we define two local search algorithms that converge monotonously to a local optimum, offering a feasible solution at any time. To demonstrate the effectiveness of these algorithms in tackling computer vision tasks, we apply them to instances of the problem that we construct from published data, using published algorithms. We report state-of-the-art application-specific accuracy in the three above-mentioned applications.

There have been remarkable improvements in the semantic labelling task in the recent years. However, the state of the art methods rely on large-scale pixel-level annotations. This paper studies the problem of training a pixel-wise semantic labeller network from image-level annotations of the present object classes. Recently, it has been shown that high quality seeds indicating discriminative object regions can be obtained from image-level labels. Without additional information, obtaining the full extent of the object is an inherently ill-posed problem due to co-occurrences. We propose using a saliency model as additional information and hereby exploit prior knowledge on the object extent and image statistics. We show how to combine both information sources in order to recover 80% of the fully supervised performance - which is the new state of the art in weakly supervised training for pixel-wise semantic labelling.

Inspired by recent advances of deep learning in instance segmentation and object tracking, we introduce the concept of convnet-based guidance applied to video object segmentation. Our model proceeds on a per-frame basis, guided by the output of the previous frame towards the object of interest in the next frame. We demonstrate that highly accurate object segmentation in videos can be enabled by using a convolutional neural network (convnet) trained with static images only. The key component of our approach is a combination of offline and online learning strategies, where the former produces a refined mask from the previous frame estimate and the latter allows to capture the appearance of the specific object instance. Our method can handle different types of input annotations such as bounding boxes and segments while leveraging an arbitrary amount of annotated frames. Therefore our system is suitable for diverse applications with different requirements in terms of accuracy and efficiency. In our extensive evaluation, we obtain competitive results on three different datasets, independently from the type of input annotation.

Social relations are the foundation of human daily life. Developing techniques to analyze such relations from visual data bears great potential to build machines that better understand us and are capable of interacting with us at a social level. Previous investigations have remained partial due to the overwhelming diversity and complexity of the topic and consequently have only focused on a handful of social relations. In this paper, we argue that the domain-based theory from social psychology is a great starting point to systematically approach this problem. The theory provides coverage of all aspects of social relations and equally is concrete and predictive about the visual attributes and behaviors defining the relations included in each domain. We provide the first dataset built on this holistic conceptualization of social life that is composed of a hierarchical label space of social domains and social relations. We also contribute the first models to recognize such domains and relations and find superior performance for attribute based features. Beyond the encouraging performance of the attribute based approach, we also find interpretable features that are in accordance with the predictions from social psychology literature. Beyond our findings, we believe that our contributions more tightly interleave visual recognition and social psychology theory that has the potential to complement the theoretical work in the area with empirical and data-driven models of social life.

Tracking multiple persons in a monocular video of a crowded scene is a challenging task. Humans can master it even if they loose track of a person locally by re-identifying the same person based on their appearance. Care must be taken across long distances, as similar-looking persons need not be identical. In this work, we propose a novel graph-based formulation that links and clusters person hypotheses over time by solving an instance of a minimum cost lifted multicut problem. Our model generalizes previous works by introducing a mechanism for adding long-range attractive connections between nodes in the graph without modifying the original set of feasible solutions. This allows us to reward tracks that assign detections of similar appearance to the same person in a way that does not introduce implausible solutions. To effectively match hypotheses over longer temporal gaps we develop new deep architectures for re-identification of people. They combine holistic representations extracted with deep networks and body pose layout obtained with a state-of-the-art pose estimation model. We demonstrate the effectiveness of our formulation by reporting a new state-of-the-art for the MOT16 benchmark. The code and pre-trained models are publicly available.

Due to the importance of zero-shot learning, the number of proposed approaches has increased steadily recently. We argue that it is time to take a step back and to analyze the status quo of the area. The purpose of this paper is three-fold. First, given the fact that there is no agreed upon zero-shot learning benchmark, we first define a new benchmark by unifying both the evaluation protocols and data splits. This is an important contribution as published results are often not comparable and sometimes even flawed due to, e.g. pre-training on zero-shot test classes. Second, we compare and analyze a significant number of the state-of-the-art methods in depth, both in the classic zero-shot setting but also in the more realistic generalized zero-shot setting. Finally, we discuss limitations of the current status of the area which can be taken as a basis for advancing it.

Convnets have enabled significant progress in pedestrian detection recently, but there are still open questions regarding suitable architectures and training data. We revisit CNN design and point out key adaptations, enabling plain FasterRCNN to obtain state-of-the-art results on the Caltech dataset. To achieve further improvement from more and better data, we introduce CityPersons, a new set of person annotations on top of the Cityscapes dataset. The diversity of CityPersons allows us for the first time to train one single CNN model that generalizes well over multiple benchmarks. Moreover, with additional training with CityPersons, we obtain top results using FasterRCNN on Caltech, improving especially for more difficult cases (heavy occlusion and small scale) and providing higher localization quality.

Object detectors have hugely profited from moving towards an end-to-end learning paradigm: proposals, features, and the classifier becoming one neural network improved results two-fold on general object detection. One indispensable component is non-maximum suppression (NMS), a post-processing algorithm responsible for merging all detections that belong to the same object. The de facto standard NMS algorithm is still fully hand-crafted, suspiciously simple, and - being based on greedy clustering with a fixed distance threshold - forces a trade-off between recall and precision. We propose a new network architecture designed to perform NMS, using only boxes and their score. We report experiments for person detection on PETS and for general object categories on the COCO dataset. Our approach shows promise providing improved localization and occlusion handling.

Users like sharing personal photos with others through social media. At the same time, they might want to make automatic identification in such photos difficult or even impossible. Classic obfuscation methods such as blurring are not only unpleasant but also not as effective as one would expect [28, 37, 18]. Recent studies on adversarial image perturbations (AIP) suggest that it is possible to confuse recognition systems effectively without unpleasant artifacts. However, in the presence of counter measures against AIPs [7], it is unclear how effective AIP would be in particular when the choice of counter measure is unknown. Game theory provides tools for studying the interaction between agents with uncertainties in the strategies. We introduce a general game theoretical framework for the user-recogniser dynamics, and present a case study that involves current state of the art AIP and person recognition techniques. We derive the optimal strategy for the user that assures an upper bound on the recognition rate independent of the recogniser's counter measure. Code is available at https://goo.gl/hgvbNK.

With an increasing number of users sharing information online, privacy implications entailing such actions are a major concern. For explicit content, such as user profile or GPS data, devices (e.g. mobile phones) as well as web services ( e.g. facebook) offer to set privacy settings in order to enforce the users' privacy preferences. We propose the first approach that extends this concept to image content in the spirit of a Visual Privacy Advisor. First, we categorize personal information in images into 68 image attributes and collect a dataset, which allows us to train models that predict such information directly from images. Second, we run a user study to understand the privacy preferences of different users w.r.t. such attributes. Third, we propose models that predict user specific privacy score from images in order to enforce the users' privacy preferences. Our model is trained to predict the user specific privacy risk and even outperforms the judgment of the users, who often fail to follow their own privacy preferences on image data.

This paper proposes the novel Pose Guided Person Generation Network (PG$^2$) that allows to synthesize person images in arbitrary poses, based on an image of that person and a novel pose. Our generation framework PG$^2$ utilizes the pose information explicitly and consists of two key stages: pose integration and image refinement. In the first stage the condition image and the target pose are fed into a U-Net-like network to generate an initial but coarse image of the person with the target pose. The second stage then refines the initial and blurry result by training a U-Net-like generator in an adversarial way. Extensive experimental results on both 128$\times$64 re-identification images and 256$\times$256 fashion photos show that our model generates high-quality person images with convincing details.

Audio description (AD) provides linguistic descriptions of movies and allows visually impaired people to follow a movie along with their peers. Such descriptions are by design mainly visual and thus naturally form an interesting data source for computer vision and computational linguistics. In this work we propose a novel dataset which contains transcribed ADs, which are temporally aligned to full length movies. In addition we also collected and aligned movie scripts used in prior work and compare the two sources of descriptions. We introduce the Large Scale Movie Description Challenge (LSMDC) which contains a parallel corpus of 128,118 sentences aligned to video clips from 200 movies (around 150 h of video in total). The goal of the challenge is to automatically generate descriptions for the movie clips. First we characterize the dataset by benchmarking different approaches for generating video descriptions. Comparing ADs to scripts, we find that ADs are more visual and describe precisely what is shown rather than what should happen according to the scripts created prior to movie production. Furthermore, we present and compare the results of several teams who participated in the challenges organized in the context of two workshops at ICCV 2015 and ECCV 2016.

Marker-less motion capture has seen great progress, but most state-of-the-art approaches fail to reliably track articulated human body motion with a very low number of cameras, let alone when applied in outdoor scenes with general background. In this paper, we propose a method for accurate marker-less capture of articulated skeleton motion of several subjects in general scenes, indoors and outdoors, even from input filmed with as few as two cameras. The new algorithm combines the strengths of a discriminative image-based joint detection method with a model-based generative motion tracking algorithm through an unified pose optimization energy. The discriminative part-based pose detection method is implemented using Convolutional Networks (ConvNet) and estimates unary potentials for each joint of a kinematic skeleton model. These unary potentials serve as the basis of a probabilistic extraction of pose constraints for tracking by using weighted sampling from a pose posterior that is guided by the model. In the final energy, we combine these constraints with an appearance-based model-to-image similarity term. Poses can be computed very efficiently using iterative local optimization, since joint detection with a trained ConvNet is fast, and since our formulation yields a combined pose estimation energy with analytic derivatives. In combination, this enables to track full articulated joint angles at state-of-the-art accuracy and temporal stability with a very low number of cameras. Our method is efficient and lends itself to implementation on parallel computing hardware, such as GPUs. We test our method extensively and show its advantages over related work on many indoor and outdoor data sets captured by ourselves, as well as data sets made available to the community by other research labs. The availability of good evaluation data sets is paramount for scientific progress, and many existing test data sets focus on controlled indoor settings, do not feature much variety in the scenes, and often lack a large corpus of data with ground truth annotation. We therefore further contribute with a new extensive test data set called MPI-MARCOnI for indoor and outdoor marker-less motion capture that features 12 scenes of varying complexity and varying camera count, and that features ground truth reference data from different modalities, ranging from manual joint annotations to marker-based motion capture results. Our new method is tested on these data, and the data set will be made available to the community.

Scaling up visual category recognition to large numbers of classes remains challenging. A promising research direction is zero-shot learning, which does not require any training data to recognize new classes, but rather relies on some form of auxiliary information describing the new classes. Ultimately, this may allow to use textbook knowledge that humans employ to learn about new classes by transferring knowledge from classes they know well. The most successful zero-shot learning approaches currently require a particular type of auxiliary information - namely attribute annotations performed by humans - that is not readily available for most classes. Our goal is to circumvent this bottleneck by substituting such annotations by extracting multiple pieces of information from multiple unstructured text sources readily available on the web. To compensate for the weaker form of auxiliary information, we incorporate stronger supervision in the form of semantic part annotations on the classes from which we transfer knowledge. We achieve our goal by a joint embedding framework that maps multiple text parts as well as multiple semantic parts into a common space. Our results consistently and significantly improve on the state-of-the-art in zero-short recognition and retrieval.

Visual understanding of complex urban street scenes is an enabling factor for a wide range of applications. Object detection has benefited enormously from large-scale datasets, especially in the context of deep learning. For semantic urban scene understanding, however, no current dataset adequately captures the complexity of real-world urban scenes. To address this, we introduce Cityscapes, a benchmark suite and large-scale dataset to train and test approaches for pixel-level and instance-level semantic labeling. Cityscapes is comprised of a large, diverse set of stereo video sequences recorded in streets from 50 different cities. 5000 of these images have high quality pixel-level annotations, 20 000 additional images have coarse annotations to enable methods that leverage large volumes of weakly-labeled data. Crucially, our effort exceeds previous attempts in terms of dataset size, annotation richness, scene variability, and complexity. Our accompanying empirical study provides an in-depth analysis of the dataset characteristics, as well as a performance evaluation of several state-of-the-art approaches based on our benchmark.

State-of-the-art learning based boundary detection methods require extensive training data. Since labelling object boundaries is one of the most expensive types of annotations, there is a need to relax the requirement to carefully annotate images to make both the training more affordable and to extend the amount of training data. In this paper we propose a technique to generate weakly supervised annotations and show that bounding box annotations alone suffice to reach high-quality object boundaries without using any object-specific boundary annotations. With the proposed weak supervision techniques we achieve the top performance on the object boundary detection task, outperforming by a large margin the current fully supervised state-of-theart methods.

In order to push the performance on realistic computer vision tasks, the number of classes in modern benchmark datasets has significantly increased in recent years. This increase in the number of classes comes along with increased ambiguity between the class labels, raising the question if top-1 error is the right performance measure. In this paper, we provide an extensive comparison and evaluation of established multiclass methods comparing their top-k performance both from a practical as well as from a theoretical perspective. Moreover, we introduce novel top-k loss functions as modifications of the softmax and the multiclass SVM losses and provide efficient optimization schemes for them. In the experiments, we compare on various datasets all of the proposed and established methods for top-k error optimization. An interesting insight of this paper is that the softmax loss yields competitive top-k performance for all k simultaneously. For a specific top-k error, our new top-k losses lead typically to further improvements while being faster to train than the softmax.

This paper considers the task of articulated human pose estimation of multiple people in real world images. We propose an approach that jointly solves the tasks of detection and pose estimation: it infers the number of persons in a scene, identifies occluded body parts, and disambiguates body parts between people in close proximity of each other. This joint formulation is in contrast to previous strategies, that address the problem by first detecting people and subsequently estimating their body pose. We propose a partitioning and labeling formulation of a set of body-part hypotheses generated with CNN-based part detectors. Our formulation, an instance of an integer linear program, implicitly performs non-maximum suppression on the set of part candidates and groups them to form configurations of body parts respecting geometric and appearance constraints. Experiments on four different datasets demonstrate state-of-the-art results for both single person and multi person pose estimation.

State-of-the-art methods for zero-shot visual recognition formulate learning as a joint embedding problem of images and side information. In these formulations the current best complement to visual features are attributes: manuallyencoded vectors describing shared characteristics among categories. Despite good performance, attributes have limitations: (1) finer-grained recognition requires commensurately more attributes, and (2) attributes do not provide a natural language interface. We propose to overcome these limitations by training neural language models from scratch, i.e. without pre-training and only consuming words and characters. Our proposed models train end-to-end to align with the fine-grained and category-specific content of images. Natural language provides a flexible and compact way of encoding only the salient visual aspects for distinguishing categories. By training on raw text, our model can do inference on raw text as well, providing humans a familiar mode both for annotation and retrieval. Our model achieves strong performance on zero-shot text-based image retrieval and significantly outperforms the attribute-based state-of-the-art for zero-shot classification on the Caltech-UCSD Birds 200-2011 dataset.

We present a novel latent embedding model for learning a compatibility function between image and class embeddings, in the context of zero-shot classification. The proposed method augments the state-of-the-art bilinear compatibility model by incorporating latent variables. Instead of learning a single bilinear map, it learns a collection of maps with the selection, of which map to use, being a latent variable for the current image-class pair. We train the model with a ranking based objective function which penalizes incorrect rankings of the true class for a given image. We empirically demonstrate that our model improves the state-of-the-art for various class embeddings consistently on three challenging publicly available datasets for the zero-shot setting. Moreover, our method leads to visually highly interpretable results with clear clusters of different fine-grained object properties that correspond to different latent variable maps.

Encouraged by the recent progress in pedestrian detection, we investigate the gap between current state-of-the-art methods and the "perfect single frame detector". We enable our analysis by creating a human baseline for pedestrian detection (over the Caltech dataset), and by manually clustering the recurrent errors of a top detector. Our results characterise both localisation and background-versusforeground errors. To address localisation errors we study the impact of training annotation noise on the detector performance, and show that we can improve even with a small portion of sanitised training data. To address background/foreground discrimination, we study convnets for pedestrian detection, and discuss which factors affect their performance. Other than our in-depth analysis, we report top performance on the Caltech dataset, and provide a new sanitised set of training and test annotations.

Automatic synthesis of realistic images from text would be interesting and useful, but current AI systems are still far from this goal. However, in recent years generic and powerful recurrent neural network architectures have been developed to learn discriminative text feature representations. Meanwhile, deep convolutional generative adversarial networks (GANs) have begun to generate highly compelling images of specific categories such as faces, album covers, room interiors and flowers. In this work, we develop a novel deep architecture and GAN formulation to effectively bridge these advances in text and image modeling, translating visual concepts from characters to pixels. We demonstrate the capability of our model to generate plausible images of birds and flowers from detailed text descriptions.

The ubiquity of portable mobile devices equipped with built-in cameras have led to a transformation in how and when digital images are captured, shared, and archived. Photographs and videos from social gatherings, public events, and even crime scenes are commonplace online. While the spontaneity afforded by these devices have led to new personal and creative outlets, privacy concerns of bystanders (and indeed, in some cases, unwilling subjects) have remained largely unaddressed. We present I-Pic, a trusted software platform that integrates digital capture with user-defined privacy. In I-Pic, users choose alevel of privacy (e.g., image capture allowed or not) based upon social context (e.g., out in public vs. with friends vs. at workplace). Privacy choices of nearby users are advertised via short-range radio, and I-Pic-compliant capture platforms generate edited media to conform to privacy choices of image subjects. I-Pic uses secure multiparty computation to ensure that users' visual features and privacy choices are not revealed publicly, regardless of whether they are the subjects of an image capture. Just as importantly, I-Pic preserves the ease-of-use and spontaneous nature of capture and sharing between trusted users. Our evaluation of I-Pic shows that a practical, energy-efficient system that conforms to the privacy choices of many users within a scene can be built and deployed using current hardware.

Generative Adversarial Networks (GANs) have recently demonstrated the capability to synthesize compelling real-world images, such as room interiors, album covers, manga, faces, birds, and flowers. While existing models can synthesize images based on global constraints such as a class label or caption, they do not provide control over pose or object location. We propose a new model, the Generative Adversarial What-Where Network (GAWWN), that synthesizes images given instructions describing what content to draw in which location. We show high-quality 128 × 128 image synthesis on the Caltech-UCSD Birds dataset, conditioned on both informal text descriptions and also object location. Our system exposes control over both the bounding box around the bird and its constituent parts. By modeling the conditional distributions over part locations, our system also enables conditioning on arbitrary subsets of parts (e.g. only the beak and tail), yielding an efficient interface for picking part locations.

We address the problem of 3D pose estimation of multiple humans from multiple views. The transition from single to multiple human pose estimation and from the 2D to 3D space is challenging due to a much larger state space, occlusions and across-view ambiguities when not knowing the identity of the humans in advance. To address these problems, we first create a reduced state space by triangulation of corresponding pairs of body parts obtained by part detectors for each camera view. In order to resolve ambiguities of wrong and mixed parts of multiple humans after triangulation and also those coming from false positive detections, we introduce a 3D pictorial structures (3DPS) model. Our model builds on multi-view unary potentials, while a prior model is integrated into pairwise and ternary potential functions. To balance the potentials' influence, the model parameters are learnt using a Structured SVM (SSVM). The model is generic and applicable to both single and multiple human pose estimation. To evaluate our model on single and multiple human pose estimation, we rely on four different datasets. We first analyse the contribution of the potentials and then compare our results with related work where we demonstrate superior performance.

Current top performing object detectors employ detection proposals to guide the search for objects, thereby avoiding exhaustive sliding window search across images. Despite the popularity and widespread use of detection proposals, it is unclear which trade-offs are made when using them during object detection. We provide an in-depth analysis of twelve proposal methods along with four baselines regarding proposal repeatability, ground truth annotation recall on PASCAL, ImageNet, and MS COCO, and their impact on DPM, R-CNN, and Fast R-CNN detection performance. Our analysis shows that for object detection improving proposal localisation accuracy is as important as improving recall. We introduce a novel metric, the average recall (AR), which rewards both high recall and good localisation and correlates surprisingly well with detection performance. Our findings show common strengths and weaknesses of existing methods, and provide insights and metrics for selecting and tuning proposal methods.

Marker-based and marker-less optical skeletal motion-capture methods use an outside-in arrangement of cameras placed around a scene, with viewpoints converging on the center. They often create discomfort with marker suits, and their recording volume is severely restricted and often constrained to indoor scenes with controlled backgrounds. Alternative suit-based systems use several inertial measurement units or an exoskeleton to capture motion with an inside-in setup, i.e. without external sensors. This makes capture independent of a confined volume, but requires substantial, often constraining, and hard to set up body instrumentation. Therefore, we propose a new method for real-time, marker-less, and egocentric motion capture: estimating the full-body skeleton pose from a lightweight stereo pair of fisheye cameras attached to a helmet or virtual reality headset - an optical inside-in method, so to speak. This allows full-body motion capture in general indoor and outdoor scenes, including crowded scenes with many people nearby, which enables reconstruction in larger-scale activities. Our approach combines the strength of a new generative pose estimation framework for fisheye views with a ConvNet-based body-part detector trained on a large new dataset. It is particularly useful in virtual reality to freely roam and interact, while seeing the fully motion-captured virtual body.

Image classification has advanced significantly in recent years with the availability of large-scale image sets. However, fine-grained classification remains a major challenge due to the annotation cost of large numbers of fine-grained categories. This project shows that compelling classification performance can be achieved on such categories even without labeled training data. Given image and class embeddings, we learn a compatibility function such that matching embeddings are assigned a higher score than mismatching ones; zero-shot classification of an image proceeds by finding the label yielding the highest joint compatibility score. We use state-of-the-art image features and focus on different supervised attributes and unsupervised output embeddings either derived from hierarchies or learned from unlabeled text corpora. We establish a substantially improved state-of-the-art on the Animals with Attributes and Caltech-UCSD Birds datasets. Most encouragingly, we demonstrate that purely unsupervised output embeddings (learned from Wikipedia and improved with finegrained text) achieve compelling results, even outperforming the previous supervised state-of-the-art. By combining different output embeddings, we further improve results.

We present a novel method for accurate marker-less capture of articulated skeleton motion of several subjects in general scenes, indoors and outdoors, even from input filmed with as few as two cameras. Our approach unites a discriminative image-based joint detection method with a model-based generative motion tracking algorithm through a combined pose optimization energy. The discriminative part-based pose detection method, implemented using Convolutional Networks (ConvNet), estimates unary potentials for each joint of a kinematic skeleton model. These unary potentials are used to probabilistically extract pose constraints for tracking by using weighted sampling from a pose posterior guided by the model. In the final energy, these constraints are combined with an appearance-based model-to-image similarity term. Poses can be computed very efficiently using iterative local optimization, as ConvNet detection is fast, and our formulation yields a combined pose estimation energy with analytic derivatives. In combination, this enables to track full articulated joint angles at state-of-the-art accuracy and temporal stability with a very low number of cameras.

In this paper we study the use of convolutional neural networks (convnets) for the task of pedestrian detection. Despite their recent diverse successes, convnets historically underperform compared to other pedestrian detectors. We deliberately omit explicitly modelling the problem into the network (e.g. parts or occlusion modelling) and show that we can reach competitive performance without bells and whistles. In a wide range of experiments we analyse small and big convnets, their architectural choices, parameters, and the influence of different training data, including pretraining on surrogate tasks. We present the best convnet detectors on the Caltech and KITTI dataset. On Caltech our convnets reach top performance both for the Caltech1x and Caltech10x training setup. Using additional data at training time our strongest convnet model is competitive even to detectors that use additional data (optical flow) at test time.

Video segmentation has become an important and active research area with a large diversity of proposed approaches. Graph-based methods, enabling top-performance on recent benchmarks, consist of three essential components: 1. powerful features account for object appearance and motion similarities; 2. spatio-temporal neighborhoods of pixels or superpixels (the graph edges) are modeled using a combination of those features; 3. video segmentation is formulated as a graph partitioning problem. While a wide variety of features have been explored and various graph partition algorithms have been proposed, there is surprisingly little research on how to construct a graph to obtain the best video segmentation performance. This is the focus of our paper. We propose to combine features by means of a classifier, use calibrated classifier outputs as edge weights and define the graph topology by edge selection. By learning the graph (without changes to the graph partitioning method), we improve the results of the best performing video segmentation algorithm by 6% on the challenging VSB100 benchmark, while reducing its runtime by 55%, as the learnt graph is much sparser.

Audio Description (AD) provides linguistic descriptions of movies and allows visually impaired people to follow a movie along with their peers. Such descriptions are by design mainly visual and thus naturally form an interesting data source for computer vision and computational linguistics. In this work we propose a novel dataset which contains transcribed ADs, which are temporally aligned to full length HD movies. In addition we also collected the aligned movie scripts which have been used in prior work and compare the two different sources of descriptions. In total the MPII Movie Description dataset (MPII-MD) contains a parallel corpus of over 68K sentences and video snippets from 94 HD movies. We characterize the dataset by benchmarking different approaches for generating video descriptions. Comparing ADs to scripts, we find that ADs are far more visual and describe precisely what is shown rather than what should happen according to the scripts created prior to movie production.

Tracking multiple targets in a video, based on a finite set of detection hypotheses, is a persistent problem in computer vision. A common strategy for tracking is to first select hypotheses spatially and then to link these over time while maintaining disjoint path constraints [14, 15, 24]. In crowded scenes multiple hypotheses will often be similar to each other making selection of optimal links an unnecessary hard optimization problem due to the sequential treatment of space and time. Embracing this observation, we propose to link and cluster plausible detections jointly across space and time. Specifically, we state multi-target tracking as a Minimum Cost Subgraph Multicut Problem. Evidence about pairs of detection hypotheses is incorporated whether the detections are in the same frame, neighboring frames or distant frames. This facilitates long-range re-identification and within-frame clustering. Results for published benchmark sequences demonstrate the superiority of this approach.

This paper starts from the observation that multiple top performing pedestrian detectors can be modelled by using an intermediate layer filtering low-level features in combination with a boosted decision forest. Based on this observation we propose a unifying framework and experimentally explore different filter families. We report extensive results enabling a systematic analysis. Using filtered channel features we obtain top performance on the challenging Caltech and KITTI datasets, while using only HOG+LUV as low-level features. When adding optical flow features we further improve detection quality and report the best known results on the Caltech dataset, reaching 93% recall at 1 FPPI.

We propose a novel algorithm for the task of supervised discriminative distance learning by nonlinearly embedding vectors into a low dimensional Euclidean space. We work in the challenging setting where supervision is with constraints on similar and dissimilar pairs while training. The proposed method is derived by an approximate kernelization of a linear Mahalanobis-like distance metric learning algorithm and can also be seen as a kernel neural network. The number of model parameters and test time evaluation complexity of the proposed method are O(dD) where D is the dimensionality of the input features and d is the dimension of the projection space -- this is in contrast to the usual kernelization methods as, unlike them, the complexity does not scale linearly with the number of training examples. We propose a stochastic gradient based learning algorithm which makes the method scalable (w.r.t. the number of training examples), while being nonlinear. We train the method with up to half a million training pairs of 4096 dimensional CNN features. We give empirical comparisons with relevant baselines on seven challenging datasets for the task of low dimensional semantic category based image retrieval.

Recognising persons in everyday photos presents major challenges (occluded faces, different clothing, locations, etc.) for machine vision. We propose a convnet based person recognition system on which we provide an in-depth analysis of informativeness of different body cues, impact of training data, and the common failure modes of the system. In addition, we discuss the limitations of existing benchmarks and propose more challenging ones. Our method is simple and is built on open source and open data, yet it improves the state of the art results on a large dataset of social media photos (PIPA).

The paradigm of multi-task learning is that one can achieve better generalization by learning tasks jointly and thus exploiting the similarity between the tasks rather than learning them independently of each other. While previously the relationship between tasks had to be user-defined in the form of an output kernel, recent approaches jointly learn the tasks and the output kernel. As the output kernel is a positive semidefinite matrix, the resulting optimization problems are not scalable in the number of tasks as an eigendecomposition is required in each step. Using the theory of positive semidefinite kernels we show in this paper that for a certain class of regularizers on the output kernel, the constraint of being positive semidefinite can be dropped as it is automatically satisfied for the relaxed problem. This leads to an unconstrained dual problem which can be solved efficiently. Experiments on several multi-task and multi-class data sets illustrate the efficacy of our approach in terms of computational efficiency as well as generalization performance.

Class ambiguity is typical in image classification problems with a large number of classes. When classes are difficult to discriminate, it makes sense to allow k guesses and evaluate classifiers based on the top-k error instead of the standard zero-one loss. We propose top-k multiclass SVM as a direct method to optimize for top-k performance. Our generalization of the well-known multiclass SVM is based on a tight convex upper bound of the top-k error. We propose a fast optimization scheme based on an efficient projection onto the top-k simplex, which is of its own interest. Experiments on five datasets show consistent improvements in top-k accuracy compared to various baselines.

As objects are inherently 3D, they have been modeled in 3D in the early days of computer vision. Due to the ambiguities arising from mapping 2D features to 3D models, 3D object representations have been neglected and 2D feature-based models are the predominant paradigm in object detection nowadays. While such models have achieved outstanding bounding box detection performance, they come with limited expressiveness, as they are clearly limited in their capability of reasoning about 3D shape or viewpoints. In this work, we bring the worlds of 3D and 2D object representations closer, by building an object detector which leverages the expressive power of 3D object representations while at the same time can be robustly matched to image evidence. To that end, we gradually extend the successful deformable part model [1] to include viewpoint information and part-level 3D geometry information, resulting in several different models with different level of expressiveness. We end up with a 3D object model, consisting of multiple object parts represented in 3D and a continuous appearance model. We experimentally verify that our models, while providing richer object hypotheses than the 2D object models, provide consistently better joint object localization and viewpoint estimation than the state-of-the-art multi-view and 3D object detectors on various benchmarks (KITTI [2] , 3D object classes [3] , Pascal3D+ [4] , Pascal VOC 2007 [5] , EPFL multi-view cars[6] ).

Detecting tools in surgical videos is an important ingredient for context-aware computer-assisted surgical systems. To this end, we present a new surgical tool detection dataset and a method for joint tool detection and pose estimation in 2d images. Our two-stage pipeline is data-driven and relaxes strong assumptions made by previous works regarding the geometry, number, and position of tools in the image. The first stage classifies each pixel based on local appearance only, while the second stage evaluates a tool-specific shape template to enforce global shape. Both local appearance and global shape are learned from training data. Our method is validated on a new surgical tool dataset of 2 476 images from neurosurgical microscopes, which is made freely available. It improves over existing datasets in size, diversity and detail of annotation. We show that our method significantly improves over competitive baselines from the computer vision field. We achieve 15% detection miss-rate at 10(-1) false positives per image (for the suction tube) over our surgical tool dataset. Results indicate that performing semantic labelling as an intermediate task is key for high quality detection.

Human pose estimation has made significant progress during the last years. However current datasets are limited in their coverage of the overall pose estimation challenges. Still these serve as the common sources to evaluate, train and compare different models on. In this paper we introduce a novel benchmark "MPII Human Pose" that makes a significant advance in terms of diversity and difficulty, a contribution that we feel is required for future developments in human body models. This comprehensive dataset was collected using an established taxonomy of over 800 human activities [1]. The collected images cover a wider variety of human activities than previous datasets including various recreational, occupational and householding activities, and capture people from a wider range of viewpoints. We provide a rich set of labels including positions of body joints, full 3D torso and head orientation, occlusion labels for joints and body parts, and activity labels. For each image we provide adjacent video frames to facilitate the use of motion information. Given these rich annotations we perform a detailed analysis of leading human pose estimation approaches and gaining insights for the success and failures of these methods.

In this work, we address the problem of 3D pose estimation of multiple humans from multiple views. This is a more challenging problem than single human 3D pose estimation due to the much larger state space, partial occlusions as well as across view ambiguities when not knowing the identity of the humans in advance. To address these problems, we first create a reduced state space by triangulation of corresponding body joints obtained from part detectors in pairs of camera views. In order to resolve the ambiguities of wrong and mixed body parts of multiple humans after triangulation and also those coming from false positive body part detections, we introduce a novel 3D pictorial structures (3DPS) model. Our model infers 3D human body configurations from our reduced state space. The 3DPS model is generic and applicable to both single and multiple human pose estimation. In order to compare to the state-of-the art, we first evaluate our method on single human 3D pose estimation on HumanEva-I [22] and KTH Multiview Football Dataset II [8] datasets. Then, we introduce and evaluate our method on two datasets for multiple human 3D pose estimation. In order to compare to the state-of-the art, we first evaluate our method on single human 3D pose estimation on HumanEva-I [22] and KTH Multiview Football Dataset II [8] datasets. Then, we introduce and evaluate our method on two datasets for multiple human 3D pose estimation.

Computational and memory costs restrict spectral techniques to rather small graphs, which is a serious limitation especially in video segmentation. In this paper, we propose the use of a reduced graph based on superpixels. In contrast to previous work, the reduced graph is reweighted such that the resulting segmentation is equivalent, under certain assumptions, to that of the full graph. We consider equivalence in terms of the normalized cut and of its spectral clustering relaxation. The proposed method reduces runtime and memory consumption and yields on par results in image and video segmentation. Further, it enables an efficient data representation and update for a new streaming video segmentation approach that also achieves state-of-the-art performance.

The underlying idea of multitask learning is that learning tasks jointly is better than learning each task individually. In particular, if only a few training examples are available for each task, sharing a jointly trained representation improves classification performance. In this paper, we propose a novel multitask learning method that learns a low-dimensional representation jointly with the corresponding classifiers, which are then able to profit from the latent inter-class correlations. Our method scales with respect to the original feature dimension and can be used with high-dimensional image descriptors such as the Fisher Vector. Furthermore, it consistently outperforms the current state of the art on the SUN397 scene classification benchmark with varying amounts of training data.

Summary form only given. Inspired by the ability of humans to interpret and understand 3D scenes nearly effortlessly, the problem of 3D scene understanding has long been advocated as the "holy grail" of computer vision. In the early days this problem was addressed in a bottom-up fashion without enabling satisfactory or reliable results for scenes of realistic complexity. In recent years there has been considerable progress on many sub-problems of the overall 3D scene understanding problem. As the performance for these sub-tasks starts to achieve remarkable performance levels, we argue that the problem to automatically infer and understand 3D scenes should be addressed again. In this talk we will - on the one hand - highlight progress on some essential components of scene understanding such as object class recognition and articulated pose estimation and tracking. On the other hand, we will also report on our current attempt towards 3D scene understanding in the particular case of traffic scene analysis.

Pictorial structure models are the de facto standard for 2D human pose estimation. Numerous refinements and improvements have been proposed such as discriminatively trained body part detectors, flexible body models, and local and global mixtures. While these techniques allow to achieve state-of-the-art performance for 2D pose estimation, they have not yet been extended to enable pose estimation in 3D. This paper thus proposes a multi-view pictorial structures model that builds on recent advances in 2D pose estimation and incorporates evidence across multiple viewpoints to allow for robust 3D pose estimation. We evaluate our multi-view pictorial structures approach on the HumanEva-I and MPII Cooking dataset. In comparison to related work for 3D pose estimation our approach achieves similar or better results while operating on single-frames only and not relying on activity specific motion models or tracking. Notably, our approach outperforms state-of-the-art for activities with more complex motions.

Despite the success of recent object class recognition systems, the long-standing problem of partial occlusion remains a major challenge, and a principled solution is yet to be found. In this paper we leave the beaten path of methods that treat occlusion as just another source of noise - instead, we include the occluder itself into the modelling, by mining distinctive, reoccurring occlusion patterns from annotated training data. These patterns are then used as training data for dedicated detectors of varying sophistication. In particular, we evaluate and compare models that range from standard object class detectors to hierarchical, part-based representations of occluder/occludee pairs. In an extensive evaluation we derive insights that can aid further developments in tackling the occlusion challenge.

In this paper we consider the challenging problem of articulated human pose estimation in still images. We observe that despite high variability of the body articulations, human motions and activities often simultaneously constrain the positions of multiple body parts. Modelling such higher order part dependencies seemingly comes at a cost of more expensive inference, which resulted in their limited use in state-of-the-art methods. In this paper we propose a model that incorporates higher order part dependencies while remaining efficient. We achieve this by defining a conditional model in which all body parts are connected a-priori, but which becomes a tractable tree-structured pictorial structures model once the image observations are available. In order to derive a set of conditioning variables we rely on the poselet-based features that have been shown to be effective for people detection but have so far found limited application for articulated human pose estimation. We demonstrate the effectiveness of our approach on three publicly available pose estimation benchmarks improving or being on-par with state of the art in each case.

Video segmentation research is currently limited by the lack of a benchmark dataset that covers the large variety of sub problems appearing in video segmentation and that is large enough to avoid over fitting. Consequently, there is little analysis of video segmentation which generalizes across subtasks, and it is not yet clear which and how video segmentation should leverage the information from the still-frames, as previously studied in image segmentation, alongside video specific information, such as temporal volume, motion and occlusion. In this work we provide such an analysis based on annotations of a large video dataset, where each video is manually segmented by multiple persons. Moreover, we introduce a new volume-based metric that includes the important aspect of temporal consistency, that can deal with segmentation hierarchies, and that reflects the tradeoff between over-segmentation and segmentation accuracy.

Typical approaches to articulated pose estimation combine spatial modelling of the human body with appearance modelling of body parts. This paper aims to push the state-of-the-art in articulated pose estimation in two ways. First we explore various types of appearance representations aiming to substantially improve the body part hypotheses. And second, we draw on and combine several recently proposed powerful ideas such as more flexible spatial models as well as image-conditioned spatial models. In a series of experiments we draw several important conclusions: (1) we show that the proposed appearance representations are complementary, (2) we demonstrate that even a basic tree-structure spatial human body model achieves state-of-the-art performance when augmented with the proper appearance representation, and (3) we show that the combination of the best performing appearance model with a flexible image-conditioned spatial model achieves the best result, significantly improving over the state of the art, on the ``Leeds Sports Poses'' and ``Parse'' benchmarks.

Humans use rich natural language to describe and communicate visual perceptions. In order to provide natural language descriptions for visual content, this paper combines two important ingredients. First, we generate a rich semantic representation of the visual content including e.g. object and activity labels. To predict the semantic representation we learn a CRF to model the relationships between different components of the visual input. And second, we propose to formulate the generation of natural language as a machine translation problem using the semantic representation as source language and the generated sentences as target language. For this we exploit the power of a parallel corpus of videos and textual descriptions and adapt statistical machine translation to translate between our two languages. We evaluate our video descriptions on the TACoS dataset, which contains video snippets aligned with sentence descriptions. Using automatic evaluation and human judgments we show significant improvements over several baseline approaches, motivated by prior work. Our translation approach also shows improvements over related work on an image description task.

People tracking in crowded real-world scenes is challenging due to frequent and long-term occlusions. Recent tracking methods obtain the image evidence from object (people) detectors, but typically use off-the-shelf detectors and treat them as black box components. In this paper we argue that for best performance one should explicitly train people detectors on failure cases of the overall tracker instead. To that end, we first propose a novel joint people detector that combines a state-of-the-art single person detector with a detector for pairs of people, which explicitly exploits common patterns of person-person occlusions across multiple viewpoints that are a frequent failure case for tracking in crowded scenes. To explicitly address remaining failure modes of the tracker we explore two methods. First, we analyze typical failures of trackers and train a detector explicitly on these cases. And second, we train the detector with the people tracker in the loop, focusing on the most common tracker failures. We show that our joint multi-person detector significantly improves both detection accuracy as well as tracker performance, improving the state-of-the-art on standard benchmarks.

Category models for objects or activities typically rely on supervised learning requiring sufficiently large training sets. Transferring knowledge from known categories to novel classes with no or only a few labels however is far less researched even though it is a common scenario. In this work, we extend transfer learning with semi-supervised learning to exploit unlabeled instances of (novel) categories with no or only a few labeled instances. Our proposed approach Propagated Semantic Transfer combines three main ingredients. First, we transfer information from known to novel categories by incorporating external knowledge, such as linguistic or expert-specified information, e.g., by a mid-level layer of semantic attributes. Second, we exploit the manifold structure of novel classes. More specifically we adapt a graph-based learning algorithm - so far only used for semi-supervised learning - to zero-shot and few-shot learning. Third, we improve the local neighborhood in such graph structures by replacing the raw feature-based representation with a mid-level object- or attribute-based representation. We evaluate our approach on three challenging datasets in two different applications, namely on Animals with Attributes and ImageNet for image classification and on MPII Composites for activity recognition. Our approach consistently outperforms state-of-the-art transfer and semi-supervised approaches on all datasets.

Following recent advances in detection, context modeling, and tracking, scene understanding has been the focus of renewed interest in computer vision research. This paper presents a novel probabilistic 3D scene model that integrates state-of-the-art multiclass object detection, object tracking and scene labeling together with geometric 3D reasoning. Our model is able to represent complex object interactions such as inter-object occlusion, physical exclusion between objects, and geometric context. Inference in this model allows us to jointly recover the 3D scene context and perform 3D multi-object tracking from a mobile observer, for objects of multiple categories, using only monocular video as input. Contrary to many other approaches, our system performs explicit occlusion reasoning and is therefore capable of tracking objects that are partially occluded for extended periods of time, or objects that have never been observed to their full extent. In addition, we show that a joint scene tracklet model for the evidence collected over multiple frames substantially improves performance. The approach is evaluated for different types of challenging onboard sequences. We first show a substantial improvement to the state of the art in 3D multipeople tracking. Moreover, a similar performance gain is achieved for multiclass 3D tracking of cars and trucks on a challenging dataset.

Geometric 3D reasoning at the level of objects has received renewed attention recently in the context of visual scene understanding. The level of geometric detail, however, is typically limited to qualitative representations or coarse boxes. This is linked to the fact that today's object class detectors are tuned toward robust 2D matching rather than accurate 3D geometry, encouraged by bounding-box-based benchmarks such as Pascal VOC. In this paper, we revisit ideas from the early days of computer vision, namely, detailed, 3D geometric object class representations for recognition. These representations can recover geometrically far more accurate object hypotheses than just bounding boxes, including continuous estimates of object pose and 3D wireframes with relative 3D positions of object parts. In combination with robust techniques for shape description and inference, we outperform state-of-the-art results in monocular 3D pose estimation. In a series of experiments, we analyze our approach in detail and demonstrate novel applications enabled by such an object class representation, such as fine-grained categorization of cars and bicycles, according to their 3D geometry, and ultrawide baseline matching.

Recent work has shown that the integration of visual information into text-based models can substantially improve model predictions, but so far only visual information extracted from static images has been used. In this paper, we consider the problem of grounding sentences describing actions in visual information extracted from videos. We present a general purpose corpus that aligns high quality videos with multiple natural language descriptions of the actions portrayed in the videos, together with an annotation of how similar the action descriptions are to each other. Experimental results demonstrate that a text-based model of similarity between actions improves substantially when combined with visual information from videos depicting the described actions.

Fine-grained categorization of object classes is receiving increased attention, since it promises to automate classification tasks that are difficult even for humans, such as the distinction between different animal species. In this paper, we consider fine-grained categorization for a different reason: following the intuition that fine-grained categories encode metric information, we aim to generate metric constraints from fine-grained cate-gory predictions, for the benefit of 3D scene-understanding. To that end, we propose two novel methods for fine-grained classification, both based on part information, as well as a new fine-grained category data set of car types. We demonstrate superior performance of our methods to state-of-the-art classifiers, and show first promising results for estimating the depth of objects from fine-grained category predictions from a monocular camera. 1

We consider the problem of detection and tracking of multiple people in crowded street scenes. State-of-the-art methods perform well in scenes with relatively few people, but are severely challenged by scenes with many subjects that partially occlude each other. This limitation is due to the fact that current people detectors fail when persons are strongly occluded. We observe that typical occlusions are due to overlaps between people and propose a people detector tailored to various occlusion levels. Instead of treating partial occlusions as distractions, we leverage the fact that person/person occlusions result in very characteristic appearance patterns that can help to improve detection results. We demonstrate the performance of our occlusion-aware person detector on a new dataset of people with controlled but severe levels of occlusion and on two challenging publicly available benchmarks outperforming single person detectors in each case.

Active learning aims to reduce the amount of labels required for classification. The main difficulty is to find a good trade-off between exploration and exploitation of the labeling process that depends - among other things - on the classification task, the distribution of the data and the employed classification scheme. In this paper, we analyze different sampling criteria including a novel density-based criteria and demonstrate the importance to combine exploration and exploitation sampling criteria. We also show that a time-varying combination of sampling criteria often improves performance. Finally, by formulating the criteria selection as a Markov decision process, we propose a novel feedback-driven framework based on reinforcement learning. Our method does not require prior information on the dataset or the sampling criteria but rather is able to adapt the sampling strategy during the learning process by experience. We evaluate our approach on three challenging object recognition datasets and show superior performance to previous active learning methods.

Current object class recognition systems typically target 2D bounding box localization, encouraged by benchmark data sets, such as Pascal VOC. While this seems suitable for the detection of individual objects, higher-level applications such as 3D scene understanding or 3D object tracking would benefit from more fine-grained object hypotheses incorporating 3D geometric information, such as viewpoints or the locations of individual parts. In this paper, we help narrowing the representational gap between the ideal input of a scene understanding system and object class detector output, by designing a detector particularly tailored towards 3D geometric reasoning. In particular, we extend the successful discriminatively trained deformable part models to include both estimates of viewpoint and 3D parts that are consistent across viewpoints. We experimentally verify that adding 3D geometric information comes at minimal performance loss w.r.t. 2D bounding box localization, but outperforms prior work in 3D viewpoint estimation and ultra-wide baseline matching.

State-of-the-art methods for human detection and pose estimation require many training samples for best performance. While large, manually collected datasets exist, the captured variations w.r.t. appearance, shape and pose are often uncontrolled thus limiting the overall performance. In order to overcome this limitation we propose a new technique to extend an existing training set that allows to explicitly control pose and shape variations. For this we build on recent advances in computer graphics to generate samples with realistic appearance and background while modifying body shape and pose. We validate the effectiveness of our approach on the task of articulated human detection and articulated pose estimation. We report close to state of the art results on the popular Image Parsing [25] human pose estimation benchmark and demonstrate superior performance for articulated human detection. In addition we define a new challenge of combined articulated human detection and pose estimation in real-world scenes.

While activity recognition is a current focus of research the challenging problem of fine-grained activity recognition is largely overlooked. We thus propose a novel database of 65 cooking activities, continuously recorded in a realistic setting. Activities are distinguished by fine-grained body motions that have low inter-class variability and high intra-class variability due to diverse subjects and ingredients. We benchmark two approaches on our dataset, one based on articulated pose tracks and the second using holistic video features. While the holistic approach outperforms the pose-based approach, our evaluation suggests that fine-grained activities are more difficult to detect and the body model can help in those cases. Providing high-resolution videos as well as an intermediate pose representation we hope to foster research in fine-grained activity recognition.

Pedestrian detection is a key problem in computer vision, with several applications that have the potential to positively impact quality of life. In recent years, the number of approaches to detecting pedestrians in monocular images has grown steadily. However, multiple data sets and widely varying evaluation protocols are used, making direct comparisons difficult. To address these shortcomings, we perform an extensive evaluation of the state of the art in a unified framework. We make three primary contributions: 1) We put together a large, well-annotated, and realistic monocular pedestrian detection data set and study the statistics of the size, position, and occlusion patterns of pedestrians in urban scenes, 2) we propose a refined per-frame evaluation methodology that allows us to carry out probing and informative comparisons, including measuring performance in relation to scale and occlusion, and 3) we evaluate the performance of sixteen pretrained state-of-the-art detectors across six data sets. Our study allows us to assess the state of the art and provides a framework for gauging future efforts. Our experiments show that despite significant progress, performance still has much room for improvement. In particular, detection is disappointing at low resolutions and for partially occluded pedestrians.

Consider the problem of recognizing an object that is partially occluded in an image. The visible portions are likely to match learned appearance models for the object, but hidden portions will not. The (hypothetical) ideal system would consider only the visible object information, correctly ignoring all occluded regions. In purely 2D recognition, this requires inferring the occlusion present, which is a significant challenge since the number of possible occlusion masks is, in principle, exponential. We simplify the problem, considering only a small subset of the most likely occlusions (top, bottom, left, and right halves) and noting that some mismatch is tolerable. We train partial-object detectors tailored exactly to each of these few cases. In addition, we reason about objects in 3D and incorporate sensed geometry, as from an RGB-depth camera, along with visual imagery. This allows explicit occlusion masks to be constructed for each object hypothesis. The masks specify how much to trust each partial template, based on their overlap with visible object regions. Only the visible evidence contributes to our object reasoning.

Robustly detecting people in real world scenes is a fundamental and challenging task in computer vision. State-of-the-art approaches use powerful learning methods and manually annotated image data. Importantly, these learning based approaches rely on the fact that the collected training data is representative of all relevant variations necessary to detect people. Rather than to collect and annotate ever more training data, this paper explores the possibility to use a 3D human shape and pose model from computer graphics to add relevant shape information to learn more powerful people detection models. By sampling from the space of 3D shapes we are able to control data variability while covering the major shape variations of humans which are often difficult to capture when collecting real-world training images. We evaluate our data generation method for a people detection model based on pictorial structures. As we show on a challenging multi-viewpoint dataset, the additional information contained in the 3D shape model helps to outperform models trained on image data alone (see e.g. Fig. 1).

People detection is an important task for a wide range of applications in computer vision. State-of-the-art methods learn appearance based models requiring tedious collection and annotation of large data corpora. Also, obtaining data sets representing all relevant variations with sufficient accuracy for the intended application domain at hand is often a non-trivial task. Therefore this paper investigates how 3D shape models from computer graphics can be leveraged to ease training data generation. In particular we employ a rendering-based reshaping method in order to generate thousands of synthetic training samples from only a few persons and views. We evaluate our data generation method for two different people detection models. Our experiments on a challenging multi-view dataset indicate that the data from as few as eleven persons suffices to achieve good performance. When we additionally combine our synthetic training samples with real data we even outperform existing state-of-the-art methods.

While knowledge transfer (KT) between object classes has been accepted as a promising route towards scalable recognition, most experimental KT studies are surprisingly limited in the number of object classes considered. To support claims of KT w.r.t. scalability we thus advocate to evaluate KT in a large-scale setting. To this end, we provide an extensive evaluation of three popular approaches to KT on a recently proposed large-scale data set, the ImageNet Large Scale Visual Recognition Competition 2010 data set. In a first setting they are directly compared to one-vs-all classification often neglected in KT papers and in a second setting we evaluate their ability to enable zero-shot learning. While none of the KT methods can improve over one-vs-all classification they prove valuable for zero-shot learning, especially hierarchical and direct similarity based KT. We also propose and describe several extensions of the evaluated approaches that are necessary for this large-scale study.

Scene understanding from a monocular, moving camera is a challenging problem with a number of applications including robotics and automotive safety. While recent systems have shown that this is best accomplished with a 3D scene model, handling of partial object occlusion is still unsatisfactory. In this paper we propose an approach that tightly integrates monocular 3D scene tracking-by-detection with explicit object-object occlusion reasoning. Full object and object part detectors are combined in a mixture of experts based on their expected visibility, which is obtained from the 3D scene model. For the difficult case of multi-people tracking, we demonstrate that our approach yields more robust detection and tracking of partially visible pedestrians, even when they are occluded over long periods of time. Our approach is evaluated on two challenging sequences recorded from a moving camera in busy pedestrian zones and outperforms several state-of-the-art approaches.

This paper considers scalable and unobtrusive activity recognition using on-body sensing for context awareness in wearable computing. Common methods for activity recognition rely on supervised learning requiring substantial amounts of labeled training data. Obtaining accurate and detailed annotations of activities is challenging, preventing the applicability of these approaches in real-world settings. This paper proposes new annotation strategies that substantially reduce the required amount of annotation. We explore two learning schemes for activity recognition that effectively leverage such sparsely labeled data together with more easily obtainable unlabeled data. Experimental results on two public data sets indicate that both approaches obtain results close to fully supervised techniques. The proposed methods are robust to the presence of erroneous labels occurring in real-world annotation data.

Recognizing 3D objects from arbitrary view points is one of the\nmost fundamental problems in computer vision. A major challenge lies\nin the transition between the 3D geometry of objects and 2D\nrepresentations that can be robustly matched to natural images. Most\napproaches thus rely on 2D natural images either as the sole source of\ntraining data for building an implicit 3D representation, or by\nenriching 3D models with natural image features.\nIn this paper, we go back to the ideas from the early days of computer\nvision, by using 3D object models as the only source of information for\nbuilding a multi-view object class detector. In particular, we use\nthese models for learning 2D shape that can be robustly matched to 2D\nnatural images. Our experiments confirm the validity of our approach,\nwhich outperforms current state-of-the-art techniques on a multi-view\ndetection data set.

Automatic recovery of 3D human pose from monocular image sequences is a challenging and important research topic with numerous applications. Although current methods are able to recover 3D pose for a single person in controlled environments, they are severely challenged by real-world scenarios, such as crowded street scenes. To address this problem, we propose a three-stage process building on a number of recent advances. The first stage obtains an initial estimate of the 2D articulation and viewpoint of the person from single frames. The second stage allows early data association across frames based on tracking-by-detection. These two stages successfully accumulate the available 2D image evidence into robust estimates of 2D limb positions over short image sequences (= tracklets). The third and final stage uses those tracklet-based estimates as robust image observations to reliably recover 3D pose. We demonstrate state-of-the-art performance on the HumanEva II benchmark, and also show the applicability of our approach to articulated 3D tracking in realistic street conditions.

This paper presents a novel mixture-of-experts framework for pedestrian classification with partial occlusion handling. The framework involves a set of component-based expert classifiers trained on features derived from intensity, depth and motion. To handle partial occlusion, we compute expert weights that are related to the degree of visibility of the associated component. This degree of visibility is determined by examining occlusion boundaries, i.e. discontinuities in depth and motion. Occlusion-dependent component weights allow to focus the combined decision of the mixture-of-experts classifier on the unoccluded body parts. In experiments on extensive real-world data sets, with both partially occluded and non-occluded pedestrians, we obtain significant performance boosts over state-of-the-art approaches by up to a factor of four in reduction of false positives at constant detection rates. The dataset is made public for benchmarking purposes.

Remarkable performance has been reported to recognize single object classes. Scalability to large numbers of classes however remains an important challenge for today's recognition methods. Several authors have promoted knowledge transfer between classes as a key ingredient to address this challenge. However, in previous work the decision which knowledge to transfer has required either manual supervision or at least a few training examples limiting the scalability of these approaches. In this work we explicitly address the question of how to automatically decide which information to transfer between classes without the need of any human intervention. For this we tap into linguistic knowledge bases to provide the semantic link between sources (what) and targets (where) of knowledge transfer. We provide a rigorous experimental evaluation of different knowledge bases and state-of-the-art techniques from Natural Language Processing which goes far beyond the limited use of language in related work. We also give insights into the applicability (why) of different knowledge sources and similarity measures for knowledge transfer.

Object recognition is challenging due to high intra-class variability caused, e.g., by articulation, viewpoint changes, and partial occlusion. Successful methods need to strike a balance between being flexible enough to model such variation and discriminative enough to detect objects in cluttered, real world scenes. Motivated by these challenges we propose a latent conditional random field (CRF) based on a flexible assembly of parts. By modeling part labels as hidden nodes and developing an EM algorithm for learning from class labels alone, this new approach enables the automatic discovery of semantically meaningful object part representations. To increase the flexibility and expressiveness of the model, we learn the pairwise structure of the underlying graphical model at the level of object part interactions. Efficient gradient-based techniques are used to estimate the structure of the domain of interest and carried forward to the multi-label or object part case. Our experiments illustrate the meaningfulness of the discovered parts and demonstrate state-of-the-art performance of the approach.

Despite impressive progress in people detection the performance on challenging datasets like Caltech Pedestrians or TUD-Brussels is still unsatisfactory. In this work we show that motion features derived from optic flow yield substantial improvements on image sequences, if implemented correctly - even in the case of low-quality video and consequently degraded flow fields. Furthermore, we introduce a new feature, self-similarity on color channels, which consistently improves detection performance both for static images and for video sequences, across different datasets. In combination with HOG, these two features outperform the state-of-the-art by up to 20%. Finally, we report two insights concerning detector evaluations, which apply to classifier-based object detection in general. First, we show that a commonly under-estimated detail of training, the number of bootstrapping rounds, has a drastic influence on the relative (and absolute) performance of different feature/classifier combinations. Second, we discuss important intricacies of detector evaluation and show that current benchmarking protocols lack crucial details, which can distort evaluations.

Person detection in complex real-world scenes is a challenging problem. State-of-the-art methods typically use supervised learning relying on significant amounts of training data to achieve good detection results. However, labeling training data is tedious, expensive, and error-prone. This paper presents a novel method to improve detection performance by supplementing real-world data with synthetically generated training data. We consider the case of detecting people in crowded scenes within an AdaBoost-framework employing Haar and Histogram-of-Oriented-Gradients (HOG) features. Our evaluations on real-world video sequences of crowded scenes with significant occlusions show that the combination of real and synthetic training data significantly improves overall detection results.

Finding injured humans is one of the primary goals of any search and rescue operation. The aim of this paper is to address the task of automatically finding people lying on the ground in images taken from the on-board camera of an unmanned aerial vehicle (UAV). In this paper we evaluate various state-of-the-art visual people detection methods in the context of vision based victim detection from an UAV. The top performing approaches in this comparison are those that rely on flexible part-based representations and discriminatively trained part detectors. We discuss their strengths and weaknesses and demonstrate that by combining multiple models we can increase the reliability of the system. We also demonstrate that the detection performance can be substantially improved by integrating the height and pitch information provided by on-board sensors. Jointly these improvements allow us to significantly boost the detection performance over the current de-facto standard, which provides a substantial step towards making autonomous victim detection for UAVs practical.

Modern mobile phones enable absolute positioning based on GPS or WiFi. However, incremental positioning based on dead reckoning is an interesting source of complementary information, e.g., for indoor positioning or for filling in reception gaps. In the literature however, reasonable dead reckoning accuracies have been reported for fixed and typically a priori known device placements only, e.g., on the hip. Therefore, this paper contributes an experimental study of several published as well as novel approaches for dead reckoning in a scenario with unconstrained placement of a device in the user's trouser pockets. Utilizing the movement of a sensor in a trouser pocket due to body motion, we estimate the user's walking direction and steps robustly for arbitrary placements in the pocket and without additional body-worn sensors. We evaluate these methods on a large dataset of 23 traces of 8 different persons, and compare different approaches.

The use of a wrist-worn sensor that is able to read nearby RFID tags and the wearer's gestures has been suggested frequently as a way to both detect the objects we interact with and to identify the interaction. Making such a prototype feasible for longer-term deployments is far from solved however, as plenty of challenges remain in the hardware, embedded algorithms, and the overall design of such a bracelet-like device. This paper presents several of the challenges that emerged during the development of a functioning prototype that is able to sense interaction data for several days. We focus in particular on RFID tag reading range optimization, efficient data logging methods, meaningful evaluation techniques, and long-term deployments.

Non-rigid object detection and articulated pose estimation are two related and challenging problems in computer vision. Numerous models have been proposed over the years and often address different special cases, such as pedestrian detection or upper body pose estimation in TV footage. This paper shows that such specialization may not be necessary, and proposes a generic approach based on the pictorial structures framework. We show that the right selection of components for both appearance and spatial modeling is crucial for general applicability and overall performance of the model. The appearance of body parts is modeled using densely sampled shape context descriptors and discriminatively trained AdaBoost classifiers. Furthermore, we interpret the normalized margin of each classifier as likelihood in a generative model. Non-Gaussian relationships between parts are represented as Gaussians in the coordinate system of the joint between parts. The marginal posterior of each part is inferred using belief propagation. We demonstrate that such a model is equally suitable for both detection and pose estimation tasks, outperforming the state of the art on three recently proposed datasets.

Pedestrian detection is a key problem in computer vision, with several applications including robotics, surveillance and automotive safety. Much of the progress of the past few years has been driven by the availability of challenging public datasets. To continue the rapid rate of innovation, we introduce the Caltech Pedestrian Dataset, which is two orders of magnitude larger than existing datasets. The dataset contains richly annotated video, recorded from a moving vehicle, with challenging images of low resolution and frequently occluded people. We propose improved evaluation metrics, demonstrating that commonly used per-window measures are flawed and can fail to predict performance on full images. We also benchmark several promising detection systems, providing an overview of state-of-the-art performance and a direct, unbiased comparison of existing methods. Finally, by analyzing common failure cases, we help identify future research directions for the field.

A variety of flexible models have been proposed to detect objects in challenging real world scenes. Motivated by some of the most successful techniques, we propose a hierarchical multi-feature representation and automatically learn flexible hierarchical object models for a wide variety of object classes. To that end we not only rely on automatic selection of relevant individual features, but go beyond previous work by automatically selecting and modeling complex, long-range feature couplings within this model. To achieve this generality and flexibility our work combines structure learning in conditional random fields and discriminative parameter learning of classifiers using hierarchical features. We adopt an efficient gradient based heuristic for model selection and carry it forward to discriminative, multidimensional selection of features and their couplings for improved detection performance. Experimentally we consistently outperform the currently leading method on all 20 classes of the PASCAL VOC 2007 challenge and achieve the best published results on 16 of 20 classes.

Various powerful people detection methods exist. Surprisingly, most approaches rely on static image features only despite the obvious potential of motion information for people detection. This paper systematically evaluates different features and classifiers in a sliding-window framework. First, our experiments indicate that incorporating motion information improves detection performance significantly. Second, the combination of multiple and complementary feature types can also help improve performance. And third, the choice of the classifier-feature combination and several implementation details are crucial to reach best performance. In contrast to many recent papers experimental results are reported for four different datasets rather than using a single one. Three of them are taken from the literature allowing for direct comparison. The fourth dataset is newly recorded using an onboard camera driving through urban environment. Consequently this dataset is more realistic and more challenging than any currently available dataset.

Object class models trained on hundreds or thousands of images have shown to enable robust detection. Transferring knowledge from such models to new object classes trained from a few or even as little as one training instance however is still in its infancy. This paper designs a shape-based model that allows to easily and explicitly transfer knowledge on three different levels: transfer of individual parts' shape and appearance information, transfer of local symmetry between parts, and transfer of part topology. Due to the factorized form of the model, knowledge can either be transferred for the complete model or just partial knowledge corresponding to certain aspects of the model. The experiments clearly demonstrate that the proposed model is competitive with the state-of-the-art and enables both full and partial knowledge transfer.

Long-term activity recognition relies on wearable sensors that log the physical actions of the wearer, so that these can be analyzed afterwards. Recent progress in this field has made it feasible to log high-resolution inertial data, resulting in increasingly large data sets. We propose the use of piecewise linear approximation techniques to facilitate this analysis. This paper presents a modified version of SWAB to approximate human inertial data as efficiently as possible, together with a matching algorithm to query for similar subsequences in large activity logs. We show that our proposed algorithms are faster on human acceleration streams than the traditional ones while being comparable in accuracy to spot similar actions, benefitting post-analysis of human activity data.

This paper illustrates both challenges and benefits found in expanding a traditional game concept to a situated environment with a distributed set of wireless sensing modules. Our pervasive game equivalent of the Whac-A-Mole game, Whac-A-Bee, retains the find-and-seek aspects of the original game while extending the location, the number of players, and the time-span in which it can be played. We discuss the obstacles met during this work, and specifically address challenges in making the game robust and flexible enough for large and long-term deployments in unknown territory.

Both detection and tracking people are challenging problems, especially in complex real world scenes that commonly involve multiple people, complicated occlusions, and cluttered or even moving backgrounds. People detectors have been shown to be able to locate pedestrians even in complex street scenes, but false positives have remained frequent. The identification of particular individuals has remained challenging as well. Tracking methods are able to find a particular individual in image sequences, but are severely challenged by real-world scenarios such as crowded street scenes. In this paper, we combine the advantages of both detection and tracking in a single framework. The approximate articulation of each person is detected in every frame based on local features that model the appearance of individual body parts. Prior knowledge on possible articulations and temporal coherency within a walking cycle are modeled using a hierarchical Gaussian process latent variable model (hGPLVM). We show how the combination of these results improves hypotheses for position and articulation of each person in several subsequent frames. We present experimental results that demonstrate how this allows to detect and track multiple people in cluttered scenes with reoccurring occlusions.

We present a novel method for the discovery and detection of visual object categories based on decompositions using topic models. The approach is capable of learning a compact and low dimensional representation for multiple visual categories from multiple view points without labeling of the training instances. The learnt object components range from local structures over line segments to global silhouette-like descriptions. This representation can be used to discover object categories in a totally unsupervised fashion. Furthermore we employ the representation as the basis for building a supervised multi-category detection system making efficient use of training examples and out-performing pure features-based representations. The proposed speed-ups make the system scale to large databases. Experiments on three databases show that the approach improves the state-of-the-art in unsupervised learning as well as supervised detection. In particular we improve the state-of-the-art on the challenging PASCALpsila06 multi-class detection tasks for several categories.

In this work we propose a novel method to recognize daily routines as a probabilistic combination of activity patterns. The use of topic models enables the automatic discovery of such patterns in a user's daily routine. We report experimental results that show the ability of the approach to model and recognize daily routines without user annotation.

Object class detection in scenes of realistic complexity remains a challenging task in computer vision. Most recent approaches focus on a single and general model for object class detection. However, in particular in the context of image sequences, it may be advantageous to adapt the general model to a more object-instance specific model in order to detect this particular object reliably within the image sequence. In this work we present a generative object model that is capable to scale from a general object class model to a more specific object-instance model. This allows to detect class instances as well as to distinguish between individual object instances reliably. We experimentally evaluate the performance of the proposed system on both still images and image sequences.

Summary form only given. We describe various approaches capable of simultaneous recognition and localization of multiple object classes using a combination of generative and discriminative methods. A first approach uses a novel hierarchical representation allows to represent individual images as well as various objects classes in a single similarity invariant model. The recognition method is based on a codebook representation where appearance clusters built from edge based features are shared among several object classes. A probabilistic model allows for reliable detection of various objects in the same image. A second approach uses a dense representation and a topic distribution model to obtain an intermediate and general representation that is shared across object categories. Combined with discriminative methods these systems show excellent performance on several object categories.

Recent work in object categorization often uses local image descriptors such as SIFT to learn and detect object categories. Such descriptors explicitly code local appearance and have shown impressive results on objects with sufficient local appearance statistics. However, many important object classes such as tools, cups and other man-made artifacts seem to require features that capture the respective shape and geometric layout of those object classes. Therefore this paper compares, on a novel data collection of 10 geometric object classes, various shape-based features with appearance-based descriptors such as SIFT. The analysis includes a direct comparison of feature statistics as well as results within standard recognition frameworks, which are partly intuitive, but sometimes surprising.

In this paper we address the problem of building object class representations based on local features and fast matching in a large database. We propose an efficient algorithm for hierarchical agglomerative clustering. We examine different agglomerative and partitional clustering strategies and compare the quality of obtained clusters. Our combination of partitional-agglomerative clustering gives significant improvement in terms of efficiency while maintaining the same quality of clusters. We also propose a method for building data structures for fast matching in high dimensional feature spaces. These improvements allow to deal with large sets of training data typically used in recognition of multiple object classes. 1

This paper proposes a novel method for integrating multiple local cues, i.e. local region detectors as well as descriptors, in the context of object detection. Rather than to fuse the outputs of several distinct classifiers in a fixed setup, our approach implements a highly adaptable integration scheme, flexibly recombining the contributions of all individual cues depending on their explanatory power for each new test image. The key idea behind our approach is to integrate the cues over an estimated top-down segmentation, which allows to quantify how much each of them contributed to the object hypothesis. By combining those contributions on a per-pixel level, our approach ensures that each cue is only used for object regions for which it is confident and that potential correlations are effectively factored out. Experimental results on several benchmark data sets show that the proposed multi-cue combination scheme significantly increases detection performance compared to any of its constituent cues alone. Moreover, it provides an interesting evaluation tool to analyze the complementarity of local feature detectors and descriptors. 1

In this paper we propose an approach capable of simultaneous recognition and localization of multiple object classes using a generative model. A novel hierarchical representation allows to represent individual images as well as various objects classes in a single, scale and rotation invariant model. The recognition method is based on a codebook representation where appearance clusters built from edge based features are shared among several object classes. A probabilistic model allows for reliable detection of various objects in the same image. The approach is highly efficient due to fast clustering and matching methods capable of dealing with millions of high dimensional features. The system shows excellent performance on several object categories over a wide range of scales, in-plane rotations, background clutter, and partial occlusions. The performance of the proposed multi-object class detection approach is competitive to state of the art approaches dedicated to a single object class recognition problem.

A wide range of methods have been proposed to detect and recognize objects. However, effective and efficient multiviewpoint detection of objects is still in its infancy, since most current approaches can only handle single viewpoints or aspects. This paper proposes a general approach for multiaspect detection of objects. As the running example for detection we use pedestrians, which add another difficulty to the problem, namely human body articulations. Global appearance changes caused by different articulations and viewpoints of pedestrians are handled in a unified manner by a generalization of the Implicit Shape Model [5]. An important property of this new approach is to share local appearance across different articulations and viewpoints, therefore requiring relatively few training samples. The effectiveness of the approach is shown and compared to previous approaches on two datasets containing pedestrians with different articulations and from multiple viewpoints.

We present a novel system for generic object class detection. In contrast to most existing systems which focus on a single viewpoint or aspect, our approach can detect object instances from arbitrary viewpoints. This is achieved by combining the Implicit Shape Model for object class detection proposed by Leibe and Schiele with the multi-view specific object recognition system of Ferrari et al. After learning single-view codebooks, these are interconnected by so-called activation links, obtained through multi-view region tracks across different training views of individual object instances. During recognition, these integrated codebooks work together to determine the location and pose of the object. Experimental results demonstrate the viability of the approach and compare it to a bank of independent single-view detectors

Natural scenes constitute a very heterogeneous stimulus class. Each semantic category contains exemplars of varying typicality. It is, therefore, an interesting question whether humans can categorize natural scenes consistently into a relatively small number of categories, such as, coasts, rivers/lakes, forests, plains, and mountains. This is particularly important for applications, such as, image retrieval systems. Only if typicality is consistently perceived across different individuals, a general image-retrieval system makes sense. In this study, we use psychophysics and computational modeling to gain a deeper understanding of scene typicality. In the first psychophysical experiment, we used a forced-choice categorization task in which each of 250 natural scenes had to be classified into one of the following five categories: coasts, rivers/lakes, forests, plains, and mountains. In the second experiment, the typicality of each scene had to be rated on a 50-point scale for each of the five categories. The psychophysical results show high consistency between participants not only in the categorization of natural scenes, but also in the typicality ratings. In order to model human perception, we then employ a computational approach that uses an intermediatesemantic modeling stepby extracting local semantic concepts, such as, rock, water, and sand. Based on the human typicality ratings, we learn a psychophysically plausible distance measure that leads to a high correlation between the computational and the human ranking of natural scenes. Interestingly, model comparisons without a semantic-modeling step correlated much less with human performance, suggesting that our model is psychophysically very plausible.

Pedestrian detection in real world scenes is a challenging problem. In recent years a variety of approaches have been proposed, and impressive results have been reported on a variety of databases. This paper systematically evaluates (1) various local shape descriptors, namely Shape Context and Local Chamfer descriptor and (2) four different interest point detectors for the detection of pedestrians. Those results are compared to the standard global Chamfer matching approach. A main result of the paper is that Shape Context trained on real edge images rather than on clean pedestrian silhouettes combined with the Hessian-Laplace detector outperforms all other tested approaches. 1

In this paper, we address the problem of detecting pedestrians in crowded real-world scenes with severe overlaps. Our basic premise is that this problem is too difficult for any type of model or feature alone. Instead, we present an algorithm that integrates evidence in multiple iterations and from different sources. The core part of our method is the combination of local and global cues via probabilistic top-down segmentation. Altogether, this approach allows examining and comparing object hypotheses with high precision down to the pixel level. Qualitative and quantitative results on a large data set confirm that our method is able to reliably detect pedestrians in crowded scenes, even when they overlap and partially occlude each other. In addition, the flexible nature of our approach allows it to operate on very small training sets.

Category detection is a lively area of research. While categorization algorithms tend to agree in using local descriptors, they differ in the choice of the classifier, with some using generative models and others discriminative approaches. This paper presents a method for object category detection which integrates a generative model with a discriminative classifier. For each object category, we generate an appearance codebook, which becomes a common vocabulary for the generative and discriminative methods. Given a query image, the generative part of the algorithm finds a set of hypotheses and estimates their support in location and scale. Then, the discriminative part verifies each hypothesis on the same codebook activations. The new algorithm exploits the strengths of both original methods, minimizing their weaknesses. Experiments on several databases show that our new approach performs better than its building blocks taken separately. Moreover, experiments on two challenging multi-scale databases show that our new algorithm outperforms previously reported results

In this paper, we compare the performance of local detectors and descriptors in the context of object class recognition. Recently, many detectors/descriptors have been evaluated in the context of matching as well as invariance to viewpoint changes (Mikolajczyk and Schmid, 2004). However, it is unclear if these results can be generalized to categorization problems, which require different properties of features. We evaluate 5 state-of-the-art scale invariant region detectors and 5 descriptors. Local features are computed for 20 object classes and clustered using hierarchical agglomerative clustering. We measure the quality of appearance clusters and location distributions using entropy as well as precision. We also measure how the clusters generalize from training set to novel test data. Our results indicate that attended SIFT descriptors (Mikolajczyk and Schmid, 2005) computed on Hessian-Laplace regions perform best. Second score is obtained by salient regions (Kadir and Brady, 2001). The results also show that these two detectors provide complementary features. The new detectors/descriptors significantly improve the performance of a state-of-the art recognition approach (Leibe, et al., 2005) in pedestrian detection task

For automatic or context-aware systems a major issue is user trust, which is to a large extent determined by system reliability. For systems based on sensor input which are inherently uncertain or even uncomplete there is little hope that they will ever be perfectly reliable. In this paper we test the hypothesis if explicitly displaying the current confidence of the system increases the usability of such systems. For the example of a context-aware mobile phone, the experiments show that displaying confidence information increases the user's trust in the system.

Most face detection algorithms locate faces by classifying the content of a detection window iterating over all positions and scales of the input image. Recent developments have accelerated this process up to real-time performance at high levels of accuracy. However, even the best of today's computational systems are far from being able to compete with the detection capabilities of the human visual system. Psychophysical experiments have shown the importance of local context in the face detection process. In this paper we investigate the role of local context for face detection algorithms. In experiments on two large data sets we find that using local context can significantly increase the number of correct detections, particularly in low resolution cases, uncommon poses or individual appearances as well as occlusions.

Object recognition has reached a level where we can identify a large number of previously seen and known objects. However, the more challenging and important task of categorizing previously unseen objects remains largely unsolved. Traditionally, contour and shape based methods are regarded most adequate for handling the generalization requirements needed for this task. Appearance based methods, on the other hand, have been successful in object identification and detection scenarios. Today little work is done to systematically compare existing methods and characterize their relative capabilities for categorizing objects. In order to compare different methods we present a new database specifically tailored to the task of object categorization. It contains high-resolution color images of 80 objects from 8 different categories, for a total of 3280 images. It is used to analyze the performance of several appearance and contour based methods. The best categorization result is obtained by an appropriate combination of different methods.

Interest point detectors are commonly employed to reduce the amount of data to be processed. The ideal interest point detector would robustly select those features which are most appropriate or salient for the application and data at hand. There is however a tradeoff between the robustness and the discriminance of the selected features. Whereas robustness in terms of repeatability is relatively well explored, the discriminance of interest points is rarely discussed. This paper formalizes the notion of saliency and evaluates three state-of-the-art interest point detectors with respect to their capability of selecting salient image features in two recognition settings.

We present the light field oracle, a novel mathematical concept for the acquisition, processing and representation of light fields. We first compute a hierarchical representation from a set of sparse image samples using a combination of wavelet transform and scattered data interpolation. The light field oracle then progressively acquires image data and selectively refines this initial representation. By comparing the actual input image to the corresponding reconstruction from the wavelet pyramid, the oracle dynamically decides on whether the new sample is needed and, if necessary, inserts it into the representation. Our incremental update scheme exploits the spatial localization of wavelets and allows for highly efficient image decomposition. Likewise, image reconstruction for rendering is computed locally in the wavelet domain and does not require a global inverse transform. The wavelet hierarchy along with fast decomposition and rendering operators constitutes a powerful mathematical framework also amenable to compression.

Combining different and complementary object models promises to increase the robustness and generality of today's computer vision algorithms. This paper introduces a new method for combining different object models by determining a configuration of the models which maximizes their mutual information. The combination scheme consequently creates a unified hypothesis from multiple object models "on the fly" without prior training. To validate the effectiveness of the proposed method, the approach is applied to the detection of faces combining the output of three different models. 1

The paper explores a view-based approach to recognize free-form objects in range images. We are using a set of local features that are easy to calculate and robust to partial occlusions. By combining those features in a multidimensional histogram, we can obtain highly discriminant classifiers without the need for segmentation. Recognition is performed using either histogram matching or a probabilistic recognition algorithm. We compare the performance of both methods in the presence of occlusions and test the system on a database of almost 2000 full-sphere views of 30 free-form objects. The system achieves a recognition accuracy above 93% on ideal images, and of 89% with 20% occlusion.

The majority of today's content based image retrieval systems rely on low-level image descriptors which limit their capability to support meaningful interactions with the users. Even though relevance feedback helps, most of the current interaction paradigms are far from the semantic representations which most people use to categorize and describe image content. Therefore we propose a concept called "vocabulary-supported image retrieval" which aims to enable the user to access an image database in a more natural way. In particular this paper develops a technique to predict the system's performance with respect to the user query. This allows the system to translate the user query into an internal query which may satisfy predefined criteria such as precision and recall rates. In addition, given the performance parameters of the system's sub-components, the feasibility and the success of the retrieval process can be evaluated beforehand and optimized dynamically online.

Objects can be represented by regions of local structure as well as dependencies between these regions. The appearance of local structure can be characterized by a vector of local features measured by local operators such as Gaussian derivatives or Gabor filters. This paper presents a technique in which the appearance of objects is represented by the joint statistics of local neighborhood operators. A probabilistic technique based on joint statistics is developed for the identification of multiple objects at arbitrary positions and orientations. Furthermore, by incorporating structural dependencies, a procedure for probabilistic localization of objects is obtained. The current recognition system runs at approximately 10 Hz on a Silicon 02. Experimental results are provided and an application using a head mounted camera is described.

This article develops an analogy between object recognition and the transmission of information through a channel based on the statistical representation of the appearances of 3D objects. This analogy provides a means to quantitatively evaluate the contribution of individual receptive field vectors, and to predict the performance of the object recognition process. Transinformation also provides a quantitative measure of the discrimination provided by each viewpoint, thus permitting the determination of the most discriminant viewpoints. As an application, the article develops an active object recognition algorithm which is able to resolve ambiguities inherent in a single-view recognition algorithm.

Describes an approach to mobile robot position estimation based on principal component analysis of laser range data. An eigenspace is constructed from the principal components of a large number of range data sets. The structure of an environment, as seen by a range sensor, is represented as a family of surfaces in this space. Subsequent range data sets from the environment project as a point in this space. Associating this point to the family of surfaces gives a set of candidate positions and orientations (poses) for the sensor. These candidate poses correspond to positions and orientations in the environment which have similar range profiles. A Kalman filter can used to select the most likely candidate pose based on coherence with small movements. The first part of this paper describes how a relatively small number of depth profiles of an environment can be used to generate a complete eigenspace. This space is used to build a representation of the range scan profiles obtained from a regular grid of positions and orientations (poses). This representation has the form of a family of surfaces (a manifold). This representation converts the problem of associating a range profile to possible positions and orientations into a table lookup. As a side benefit, the method provides a simple means to detect obstacles in a range profile. The final section of the paper reviews the use of estimation theory to determine the correct pose hypothesis by tracking.

This paper describes a probabilistic object recognition technique which does not require correspondence matching of images. This technique is an extension of our earlier work (1996) on object recognition using matching of multi-dimensional receptive field histograms. In the earlier paper we have shown that multi-dimensional receptive field histograms can be matched to provide object recognition which is robust in the face of changes in viewing position and independent of image plane rotation and scale. In this paper we extend this method to compute the probability of the presence of an object in an image. The paper begins with a review of the method and previously presented experimental results. We then extend the method for histogram matching to obtain a genuine probability of the presence of an object. We present experimental results on a database of 100 objects showing that the approach is capable recognizing all objects correctly by using only a small portion of the image. Our results show that receptive field histograms provide a technique for object recognition which is robust, has low computational cost and a computational complexity which is linear with the number of pixels.

The authors (1996) introduced the use of multidimensional receptive field histograms for probabilistic object recognition. In this paper we reverse the object recognition problem by asking the question "where should we look?", when we want to verify the presence of an object, to track an object or to actively explore a scene. This paper describes the statistical framework from which we obtain a network of salient points for an object. This network of salient points may be used for fixation control in the context of active object recognition.

A mobile robot requires perception of its local environment for both sensor based locomotion and for position estimation. Occupancy grids, based on ultrasonic range data, provide a robust description of the local environment for locomotion. Unfortunately, current techniques for position estimation based on occupancy grids are both unreliable and computationally expensive. This paper reports on experiments with four techniques for position estimation using occupancy grids. A world modeling technique based on combining global and local occupancy grids is described. Techniques are described for extracting line segments from an occupancy grid based on a Hough transform. The use of an extended Kalman filter for position estimation is then adapted to this framework. Four matching techniques are presented for obtaining the innovation vector required by the Kalman filter equations. Experimental results show that matching of segments extracted from the both the local and global occupancy grids gives results which are superior to a direct matching of grids, or to a mixed matching of segments to grids.>