Unsupervised panoptic segmentation aims to partition an image into semantically meaningful regions and distinct object instances without training on manually annotated data. In contrast to prior work on unsupervised panoptic scene understanding, we eliminate the need for object-centric training data, enabling the unsupervised understanding of complex scenes. To that end, we present the first unsupervised panoptic method that directly trains on scene-centric imagery. In particular, we propose an approach to obtain high-resolution panoptic pseudo labels on complex scene-centric data, combining visual representations, depth, and motion cues. Utilizing both pseudo-label training and a panoptic self-training strategy yields a novel approach that accurately predicts panoptic segmentation of complex scenes without requiring any human annotations. Our approach significantly improves panoptic quality, e.g., surpassing the recent state of the art in unsupervised panoptic segmentation on Cityscapes by 9.4 % points in PQ.

Object-centric learning aims to decompose an input image into a set of meaningful object files (slots). These latent object representations enable a variety of downstream tasks. Yet, object-centric learning struggles on real-world datasets, which contain multiple objects of complex textures and shapes in natural everyday scenes. To address this, we introduce Guided Latent Slot Diffusion (GLASS), a novel slot-attention model that learns in the space of generated images and uses semantic and instance guidance modules to learn better slot embeddings for various downstream tasks. Our experiments show that GLASS surpasses state-of-the-art slot-attention methods by a wide margin on tasks such as (zero-shot) object discovery and conditional image generation for real-world scenes. Moreover, GLASS enables the first application of slot attention to the compositional generation of complex, realistic scenes. *

Semantic scene completion (SSC) aims to infer both the 3D geometry and semantics of a scene from single images. In contrast to prior work on SSC that heavily relies on expensive ground-truth annotations, we approach SSC in an unsupervised setting. Our novel method, SceneDINO, adapts techniques from self-supervised representation learning and 2D unsupervised scene understanding to SSC. Our training exclusively utilizes multi-view consistency self-supervision without any form of semantic or geometric ground truth. Given a single input image, SceneDINO infers the 3D geometry and expressive 3D DINO features in a feed-forward manner. Through a novel 3D feature distillation approach, we obtain unsupervised 3D semantics. In both 3D and 2D unsupervised scene understanding, SceneDINO reaches state-of-the-art segmentation accuracy. Linear probing our 3D features matches the segmentation accuracy of a current supervised SSC approach. Additionally, we showcase the domain generalization and multi-view consistency of SceneDINO, taking the first steps towards a strong foundation for single image 3D scene understanding.

Cost volumes are used in every modern optical flow estimator, but due to their computational and space complexity, they are often a limiting factor regarding both processing speed and the resolution of input frames. Motivated by our empirical observation that cost volumes lose their importance once all other network parts of, e.g., a RAFT-based pipeline have been sufficiently trained, we introduce a training strategy that allows removing the cost volume from optical flow estimators throughout training. This leads to significantly improved inference speed and reduced memory requirements. Using our training strategy, we create three different models covering different compute budgets. Our most accurate model reaches state-of-the-art accuracy while being $1.2\times$ faster and having a $6\times$ lower memory footprint than comparable models; our fastest model is capable of processing Full HD frames at $20\,\mathrm{FPS}$ using only $500\,\mathrm{MB}$ of GPU memory.

Visual relation detection (VRD) is the task of identifying the relationships between objects in a scene. VRD models trained solely on relation detection data struggle to generalize beyond the relations on which they are trained. While prompt tuning has been used to adapt vision-language models (VLMs) for VRD, it uses handcrafted prompts and struggles with novel or complex relations. We argue that instruction tuning offers a more effective solution by fine-tuning VLMs on diverse instructional data. We thus introduce ART, an Adaptive Relation Tuning framework that adapts VLMs for VRD through instruction tuning and strategic instance selection. By converting VRD datasets into an instruction tuning format and employing an adaptive sampling algorithm, ART directs the VLM to focus on informative relations while maintaining generalizability. Specifically, we focus on the relation classification, where subject-object boxes are given and the model predicts the predicate between them. We tune on a held-in set and evaluate across multiple held-out datasets of varying complexity. Our approach strongly improves over its baselines and can infer unseen relation concepts, a capability absent in mainstream VRD methods. We demonstrate ART's practical value by using the predicted relations for segmenting complex scenes.

To ensure the reliability of deep models in real-world applications, out-of-distribution (OOD) detection methods aim to distinguish samples close to the training distribution (in-distribution, ID) from those farther away (OOD). In this work, we propose a novel OOD detection method that utilizes singular value decomposition of the weight matrix of the classification head to decompose the model's activations into decisive and insignificant components, which contribute maximally, respectively minimally, to the final classifier output. We find that the subspace of insignificant components more effectively distinguishes ID from OOD data than raw activations in regimes of large distribution shifts (Far-OOD). This occurs because the classification objective leaves the insignificant subspace largely unaffected, yielding features that are ''untainted'' by the target classification task. Conversely, in regimes of smaller distribution shifts (Near-OOD), we find that activation shaping methods profit from only considering the decisive subspace, as the insignificant component can cause interference in the activation space. By combining two findings into a single approach, termed ActSub, we achieve state-of-the-art results in various standard OOD benchmarks.

Omnidirectional depth perception is essential for mobile robotics applications that require scene understanding across a full 360° field of view. Camera-based setups offer a cost-effective option by using stereo depth estimation to generate dense, high-resolution depth maps without relying on expensive active sensing. However, existing omnidirectional stereo matching approaches achieve only limited depth accuracy across diverse environments, depth ranges, and lighting conditions, due to the scarcity of real-world data. We present DFI-OmniStereo, a novel omnidirectional stereo matching method that leverages a large-scale pre-trained foundation model for relative monocular depth estimation within an iterative optimization-based stereo matching architecture. We introduce a dedicated two-stage training strategy to utilize the relative monocular depth features for our omnidirectional stereo matching before scale-invariant fine-tuning. DFI-OmniStereo achieves state-of-the-art results on the real-world Helvipad dataset, reducing disparity MAE by approximately 16% compared to the previous best omnidirectional stereo method.

Adapters provide an efficient and lightweight mechanism for adapting trained transformer models to a variety of dif-ferent tasks. However, they have often been found to be outperformed by other adaptation mechanisms including low-rank adaptation. In this paper, we provide an in-depth study of adapters, their internal structure, as well as vari-ous implementation choices. We uncover pitfalls for using adapters and suggest a concrete, improved adapter architecture, called Adapter+, that not only outperforms previous adapter implementations but surpasses a number of other, more complex adaptation mechanisms in several challenging settings. Despite this, our suggested adapter is highly robust and, unlike previous work, requires little to no manual inter-vention when addressing a novel scenario. Adapter+ reaches state-of-the-art average accuracy on the VTAB benchmark, even without a per-task hyperparameter optimization.††Code is available at https://github.com/visinf/adapter_plus.

Attribution maps are one of the most established tools to explain the functioning of computer vision models. They assign importance scores to input features, indicating how relevant each feature is for the prediction of a deep neural network. While much research has gone into proposing new attribution methods, their proper evaluation remains a difficult challenge. In this work, we propose a novel evaluation protocol that overcomes two fundamental limitations of the widely used incremental-deletion protocol, i.e., the out-of-domain issue and lacking inter-model comparisons. This allows us to evaluate 23 attribution methods and how different design choices of popular vision backbones affect their attribution quality. We find that intrinsically explainable models outperform standard models and that raw attribution values exhibit a higher attribution quality than what is known from previous work. Further, we show consistent changes in the attribution quality when varying the network design, indicating that some standard design choices promote attribution quality.

The field of explainable artificial intelligence (XAI) aims to uncover the inner workings of complex deep neural models. While being crucial for safety-critical domains, XAI inherently lacks ground-truth explanations, making its automatic evaluation an unsolved problem. We address this challenge by proposing a novel synthetic vision dataset, named FunnyBirds, and accompanying automatic evaluation protocols. Our dataset allows performing semantically meaningful image interventions, e.g., removing individual object parts, which has three important implications. First, it enables analyzing explanations on a part level, which is closer to human comprehension than existing methods that evaluate on a pixel level. Second, by comparing the model output for inputs with removed parts, we can estimate ground-truth part importances that should be reflected in the explanations. Third, by mapping individual explanations into a common space of part importances, we can analyze a variety of different explanation types in a single common framework. Using our tools, we report results for 24 different combinations of neural models and XAI methods, demonstrating the strengths and weaknesses of the assessed methods in a fully automatic and systematic manner.

Recent years have seen a growing interest in Scene Graph Generation (SGG), a comprehensive visual scene understanding task that aims to predict entity relationships using a relation encoder-decoder pipeline stacked on top of an object encoder-decoder backbone. Unfortunately, current SGG methods suffer from an information loss regarding the entities’ local-level cues during the relation encoding process. To mitigate this, we introduce the Vision rElation TransfOrmer (VETO), consisting of a novel local-level entity relation encoder. We further observe that many existing SGG methods claim to be unbiased, but are still biased towards either head or tail classes. To overcome this bias, we introduce a Mutually Exclusive ExperT (MEET) learning strategy that captures important relation features without bias towards head or tail classes. Experimental results on the VG and GQA datasets demonstrate that VETO + MEET boosts the predictive performance by up to 47% over the state of the art while being ∼ 10× smaller.1

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.

We propose an approach to domain adaptation for semantic segmentation that is both practical and highly accurate. In contrast to previous work, we abandon the use of computationally involved adversarial objectives, network ensembles and style transfer. Instead, we employ standard data augmentation techniques – photometric noise, flipping and scaling – and ensure consistency of the semantic predictions across these image transformations. We develop this principle in a lightweight self-supervised framework trained on co-evolving pseudo labels without the need for cumbersome extra training rounds. Simple in training from a practitioner’s standpoint, our approach is remarkably effective. We achieve significant improvements of the state-of-the-art segmentation accuracy after adaptation, consistent both across different choices of the backbone architecture and adaptation scenarios.

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.

Autoregressive models are a class of exact inference approaches with highly flexible functional forms, yielding state-of-the-art density estimates for natural images. Yet, the sequential ordering on the dimensions makes these models computationally expensive and limits their applicability to low-resolution imagery. In this work, we propose PixelPyramids,1a block-autoregressive approach employing a lossless pyramid decomposition with scale-specific representations to encode the joint distribution of image pixels. Crucially, it affords a sparser dependency structure compared to fully autoregressive approaches. Our PixelPyramids yield state-of-the-art results for density estimation on various image datasets, especially for high-resolution data. For CelebA-HQ 1024 × 1024, we observe that the density estimates (in terms of bits/dim) are improved to ~44% of the baseline despite sampling speeds superior even to easily parallelizable flow-based models.

We present a novel approach to unsupervised learning for video object segmentation (VOS). Unlike previous work, our formulation allows to learn dense feature representations directly in a fully convolutional regime. We rely on uniform grid sampling to extract a set of anchors and train our model to disambiguate between them on both inter- and intra-video levels. However, a naive scheme to train such a model results in a degenerate solution. We propose to prevent this with a simple regularisation scheme, accommodating the equivariance property of the segmentation task to similarity transformations. Our training objective admits efficient implementation and exhibits fast training convergence. On established VOS benchmarks, our approach exceeds the segmentation accuracy of previous work despite using significantly less training data and compute power.

Mitigating the dependence on spurious correlations present in the training dataset is a quickly emerging and important topic of deep learning. Recent approaches include priors on the feature attribution of a deep neural network (DNN) into the training process to reduce the dependence on unwanted features. However, until now one needed to trade off high-quality attributions, satisfying desirable axioms, against the time required to compute them. This in turn either led to long training times or ineffective attribution priors. In this work, we break this trade-off by considering a special class of efficiently axiomatically attributable DNNs for which an axiomatic feature attribution can be computed with only a single forward/backward pass. We formally prove that nonnegatively homogeneous DNNs, here termed $\mathcal{X}$-DNNs, are efficiently axiomatically attributable and show that they can be effortlessly constructed from a wide range of regular DNNs by simply removing the bias term of each layer. Various experiments demonstrate the advantages of $\mathcal{X}$-DNNs, beating state-of-the-art generic attribution methods on regular DNNs for training with attribution priors.

Monocular depth prediction is ill-posed by nature; hence successful approaches need to exploit the available cues to the fullest. Yet, real-world training data with depth ground-truth suffers from limited variability and data acquired from depth sensors is also sparse and prone to noise. While available datasets with semantic annotations might help to better exploit semantic cues, they are not immediately usable for depth prediction. We show how to leverage panoptic segmentation maps to boost monocular depth predictors in stereo training setups. In particular, we augment a self-supervised training scheme through panoptic-guided smoothing, panoptic-guided alignment, and panoptic left-right consistency from ground truth or inferred panoptic segmentation maps. Our approach incurs only a minor overhead, can easily be applied to a wide range of depth estimation methods that are trained at least partially using stereo pairs, providing a substantial boost in accuracy.

Abstract Standardized benchmarks have been crucial in pushing the performance of computer vision algorithms, especially since the advent of deep learning. Although leaderboards should not be over-claimed, they often provide the most objective measure of performance and are therefore important guides for research. We present MOTChallenge , a benchmark for single-camera Multiple Object Tracking (MOT) launched in late 2014, to collect existing and new data and create a framework for the standardized evaluation of multiple object tracking methods. The benchmark is focused on multiple people tracking, since pedestrians are by far the most studied object in the tracking community, with applications ranging from robot navigation to self-driving cars. This paper collects the first three releases of the benchmark: (i) MOT15 , along with numerous state-of-the-art results that were submitted in the last years, (ii) MOT16 , which contains new challenging videos, and (iii) MOT17 , that extends MOT16 sequences with more precise labels and evaluates tracking performance on three different object detectors. The second and third release not only offers a significant increase in the number of labeled boxes, but also provide labels for multiple object classes beside pedestrians, as well as the level of visibility for every single object of interest. We finally provide a categorization of state-of-the-art trackers and a broad error analysis. This will help newcomers understand the related work and research trends in the MOT community, and hopefully shed some light into potential future research directions.

Recent years have seen a rapid growth in new approaches improving the accuracy of semantic segmentation in a weakly supervised setting, i.e. with only image-level labels available for training. However, this has come at the cost of increased model complexity and sophisticated multi-stage training procedures. This is in contrast to earlier work that used only a single stage -- training one segmentation network on image labels -- which was abandoned due to inferior segmentation accuracy. In this work, we first define three desirable properties of a weakly supervised method: local consistency, semantic fidelity, and completeness. Using these properties as guidelines, we then develop a segmentation-based network model and a self-supervised training scheme to train for semantic masks from image-level annotations in a single stage. We show that despite its simplicity, our method achieves results that are competitive with significantly more complex pipelines, substantially outperforming earlier single-stage methods.

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.

Scene flow estimation has been receiving increasing attention for 3D environment perception. Monocular scene flow estimation - obtaining 3D structure and 3D motion from two temporally consecutive images - is a highly ill-posed problem, and practical solutions are lacking to date. We propose a novel monocular scene flow method that yields competitive accuracy and real-time performance. By taking an inverse problem view, we design a single convolutional neural network (CNN) that successfully estimates depth and 3D motion simultaneously from a classical optical flow cost volume. We adopt self-supervised learning with 3D loss functions and occlusion reasoning to leverage unlabeled data. We validate our design choices, including the proxy loss and augmentation setup. Our model achieves state-of-the-art accuracy among unsupervised/self-supervised learning approaches to monocular scene flow, and yields competitive results for the optical flow and monocular depth estimation sub-tasks. Semi-supervised fine-tuning further improves the accuracy and yields promising results in real-time.

Encoder-decoder networks have found widespread use in various dense prediction tasks. However, the strong reduction of spatial resolution in the encoder leads to a loss of location information as well as boundary artifacts. To address this, image-adaptive post-processing methods have shown beneficial by leveraging the high-resolution input image(s) as guidance data. We extend such approaches by considering an important orthogonal source of information: the network's confidence in its own predictions. We introduce probabilistic pixel-adaptive convolutions (PPACs), which not only depend on image guidance data for filtering, but also respect the reliability of per-pixel predictions. As such, PPACs allow for image-adaptive smoothing and simultaneously propagating pixels of high confidence into less reliable regions, while respecting object boundaries. We demonstrate their utility in refinement networks for optical flow and semantic segmentation, where PPACs lead to a clear reduction in boundary artifacts. Moreover, our proposed refinement step is able to substantially improve the accuracy on various widely used benchmarks.

General-purpose planning algorithms for automated driving combine mission, behavior, and local motion planning. Such planning algorithms map features of the environment and driving kinematics into complex reward functions. To achieve this, planning experts often rely on linear reward functions. The specification and tuning of these reward functions is a tedious process and requires significant experience. Moreover, a manually designed linear reward function does not generalize across different driving situations. In this work, we propose a deep learning approach based on inverse reinforcement learning that generates situation-dependent reward functions. Our neural network provides a mapping between features and actions of sampled driving policies of a model-predictive control-based planner and predicts reward functions for upcoming planning cycles. In our evaluation, we compare the driving style of reward functions predicted by our deep network against clustered and linear reward functions. Our proposed deep learning approach outperforms clustered linear reward functions and is at par with linear reward functions with a-priori knowledge about the situation.

General-purpose trajectory planning algorithms for automated driving utilize complex reward functions to perform a combined optimization of strategic, behavioral, and kinematic features. The specification and tuning of a single reward function is a tedious task and does not generalize over a large set of traffic situations. Deep learning approaches based on path integral inverse reinforcement learning have been successfully applied to predict local situation-dependent reward functions using features of a set of sampled driving policies. Sample-based trajectory planning algorithms are able to approximate a spatio-temporal subspace of feasible driving policies that can be used to encode the context of a situation. However, the interaction with dynamic objects requires an extended planning horizon, which depends on sequential context modeling. In this work, we are concerned with the sequential reward prediction over an extended time horizon. We present a neural network architecture that uses a policy attention mechanism to generate a low-dimensional context vector by concentrating on trajectories with a human-like driving style. Apart from this, we propose a temporal attention mechanism to identify context switches and allow for stable adaptation of rewards. We evaluate our results on complex simulated driving situations, including other moving vehicles. Our evaluation shows that our policy attention mechanism learns to focus on collision-free policies in the configuration space. Furthermore, the temporal attention mechanism learns persistent interaction with other vehicles over an extended planning horizon.

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.

Diverse image captioning models aim to learn one-to-many mappings that are innate to cross-domain datasets, such as of images and texts. Current methods for this task are based on generative latent variable models, eg. VAEs with structured latent spaces. Yet, the amount of multimodality captured by prior work is limited to that of the paired training data -- the true diversity of the underlying generative process is not fully captured. To address this limitation, we leverage the contextual descriptions in the dataset that explain similar contexts in different visual scenes. To this end, we introduce a novel factorization of the latent space, termed context-object split, to model diversity in contextual descriptions across images and texts within the dataset. Our framework not only enables diverse captioning through context-based pseudo supervision, but extends this to images with novel objects and without paired captions in the training data. We evaluate our COS-CVAE approach on the standard COCO dataset and on the held-out COCO dataset consisting of images with novel objects, showing significant gains in accuracy and diversity.

Most approaches to visual scene analysis have emphasised parallel processing of the image elements. However, one area in which the sequential nature of vision is apparent, is that of segmenting multiple, potentially similar and partially occluded objects in a scene. In this work, we revisit the recurrent formulation of this challenging problem in the context of reinforcement learning. Motivated by the limitations of the global max-matching assignment of the ground-truth segments to the recurrent states, we develop an actor-critic approach in which the actor recurrently predicts one instance mask at a time and utilises the gradient from a concurrently trained critic network. We formulate the state, action, and the reward such as to let the critic model long-term effects of the current prediction and in- corporate this information into the gradient signal. Furthermore, to enable effective exploration in the inherently high-dimensional action space of instance masks, we learn a compact representation using a conditional variational auto-encoder. We show that our actor-critic model consistently provides accuracy benefits over the recurrent baseline on standard instance segmentation benchmarks.

Deep learning approaches to optical flow estimation have seen rapid progress over the recent years. One common trait of many networks is that they refine an initial flow estimate either through multiple stages or across the levels of a coarse-to-fine representation. While leading to more accurate results, the downside of this is an increased number of parameters. Taking inspiration from both classical energy minimization approaches as well as residual networks, we propose an iterative residual refinement (IRR) scheme based on weight sharing that can be combined with several backbone networks. It reduces the number of parameters, improves the accuracy, or even achieves both. Moreover, we show that integrating occlusion prediction and bi-directional flow estimation into our IRR scheme can further boost the accuracy. Our full network achieves state-of-the-art results for both optical flow and occlusion estimation across several standard datasets.

Behavior and motion planning play an important role in automated driving. Traditionally, behavior planners instruct local motion planners with predefined behaviors. Due to the high scene complexity in urban environments, unpredictable situations may occur in which behavior planners fail to match predefined behavior templates. Recently, general-purpose planners have been introduced, combining behavior and local motion planning. These general-purpose planners allow behavior-aware motion planning given a single reward function. However, two challenges arise: First, this function has to map a complex feature space into rewards. Second, the reward function has to be manually tuned by an expert. Manually tuning this reward function becomes a tedious task. In this paper, we propose an approach that relies on human driving demonstrations to automatically tune reward functions. This study offers important insights into the driving style optimization of general-purpose planners with maximum entropy inverse reinforcement learning. We evaluate our approach based on the expected value difference between learned and demonstrated policies. Furthermore, we compare the similarity of human driven trajectories with optimal policies of our planner under learned and expert-tuned reward functions. Our experiments show that we are able to learn reward functions exceeding the level of manual expert tuning without prior domain knowledge.

In the era of end-to-end deep learning, many advances in computer vision are driven by large amounts of labeled data. In the optical flow setting, however, obtaining dense per-pixel ground truth for real scenes is difficult and thus such data is rare. Therefore, recent end-to-end convolutional networks for optical flow rely on synthetic datasets for supervision, but the domain mismatch between training and test scenarios continues to be a challenge. Inspired by classical energy-based optical flow methods, we design an unsupervised loss based on occlusion-aware bidirectional flow estimation and the robust census transform to circumvent the need for ground truth flow. On the KITTI benchmarks, our unsupervised approach outperforms previous unsupervised deep networks by a large margin, and is even more accurate than similar supervised methods trained on synthetic datasets alone. By optionally fine-tuning on the KITTI training data, our method achieves competitive optical flow accuracy on the KITTI 2012 and 2015 benchmarks, thus in addition enabling generic pre-training of supervised networks for datasets with limited amounts of ground truth.

Even though probabilistic treatments of neural networks have a long history, they have not found widespread use in practice. Sampling approaches are often too slow already for simple networks. The size of the inputs and the depth of typical CNN architectures in computer vision only compound this problem. Uncertainty in neural networks has thus been largely ignored in practice, despite the fact that it may provide important information about the reliability of predictions and the inner workings of the network. In this paper, we introduce two lightweight approaches to making supervised learning with probabilistic deep networks practical: First, we suggest probabilistic output layers for classification and regression that require only minimal changes to existing networks. Second, we employ assumed density filtering and show that activation uncertainties can be propagated in a practical fashion through the entire network, again with minor changes. Both probabilistic networks retain the predictive power of the deterministic counterpart, but yield uncertainties that correlate well with the empirical error induced by their predictions. Moreover, the robustness to adversarial examples is significantly increased.

Variational inference has experienced a recent surge in popularity owing to stochastic approaches, which have yielded practical tools for a wide range of model classes. A key benefit is that stochastic variational inference obviates the tedious process of deriving analytical expressions for closed-form variable updates. Instead, one simply needs to derive the gradient of the log-posterior, which is often much easier. Yet for certain model classes, the log-posterior itself is difficult to optimize using standard gradient techniques. One such example are random field models, where optimization based on gradient linearization has proven popular, since it speeds up convergence significantly and can avoid poor local optima. In this paper we propose stochastic variational inference with gradient linearization (SVIGL). It is similarly convenient as standard stochastic variational inference - all that is required is a local linearization of the energy gradient. Its benefit over stochastic variational inference with conventional gradient methods is a clear improvement in convergence speed, while yielding comparable or even better variational approximations in terms of KL divergence. We demonstrate the benefits of SVIGL in three applications: Optical flow estimation, Poisson-Gaussian denoising, and 3D surface reconstruction.

In this paper, we develop novel, efficient 2D encodings for 3D geometry, which enable reconstructing full 3D shapes from a single image at high resolution. The key idea is to pose 3D shape reconstruction as a 2D prediction problem. To that end, we first develop a simple baseline network that predicts entire voxel tubes at each pixel of a reference view. By leveraging well-proven architectures for 2D pixel-prediction tasks, we attain state-of-the-art results, clearly outperforming purely voxel-based approaches. We scale this baseline to higher resolutions by proposing a memory-efficient shape encoding, which recursively decomposes a 3D shape into nested shape layers, similar to the pieces of a Matryoshka doll. This allows reconstructing highly detailed shapes with complex topology, as demonstrated in extensive experiments; we clearly outperform previous octree-based approaches despite having a much simpler architecture using standard network components. Our Matryoshka networks further enable reconstructing shapes from IDs or shape similarity, as well as shape sampling.

Most convolutional neural networks use some method for gradually downscaling the size of the hidden layers. This is commonly referred to as pooling, and is applied to reduce the number of parameters, improve invariance to certain distortions, and increase the receptive field size. Since pooling by nature is a lossy process, it is crucial that each such layer maintains the portion of the activations that is most important for the network's discriminability. Yet, simple maximization or averaging over blocks, max or average pooling, or plain downsampling in the form of strided convolutions are the standard. In this paper, we aim to leverage recent results on image downscaling for the purposes of deep learning. Inspired by the human visual system, which focuses on local spatial changes, we propose detail-preserving pooling (DPP), an adaptive pooling method that magnifies spatial changes and preserves important structural detail. Importantly, its parameters can be learned jointly with the rest of the network. We analyze some of its theoretical properties and show its empirical benefits on several datasets and networks, where DPP consistently outperforms previous pooling approaches.

Teresa Botschen, Iryna Gurevych, Jan-Christoph Klie, Hatem Mousselly-Sergieh, Stefan Roth. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.

Non-local methods exploiting the self-similarity of natural signals have been well studied, for example in image analysis and restoration. Existing approaches, however, rely on k-nearest neighbors (KNN) matching in a fixed feature space. The main hurdle in optimizing this feature space w.r.t. application performance is the non-differentiability of the KNN selection rule. To overcome this, we propose a continuous deterministic relaxation of KNN selection that maintains differentiability w.r.t. pairwise distances, but retains the original KNN as the limit of a temperature parameter approaching zero. To exploit our relaxation, we propose the neural nearest neighbors block (N3 block), a novel non-local processing layer that leverages the principle of self-similarity and can be used as building block in modern neural network architectures. We show its effectiveness for the set reasoning task of correspondence classification as well as for image restoration, including image denoising and single image super-resolution, where we outperform strong convolutional neural network (CNN) baselines and recent non-local models that rely on KNN selection in hand-chosen features spaces.

We present a novel approach to noise-blind deblurring, the problem of deblurring an image with known blur, but unknown noise level. We introduce an efficient and robust solution based on a Bayesian framework using a smooth generalization of the 0-1 loss. A novel bound allows the calculation of very high-dimensional integrals in closed form. It avoids the degeneracy of Maximum a-Posteriori (MAP) estimates and leads to an effective noise-adaptive scheme. Moreover, we drastically accelerate our algorithm by using Majorization Minimization (MM) without introducing any approximation or boundary artifacts. We further speed up convergence by turning our algorithm into a neural network termed GradNet, which is highly parallelizable and can be efficiently trained. We demonstrate that our noise-blind formulation can be integrated with different priors and significantly improves existing deblurring algorithms in the noise-blind and in the known-noise case. Furthermore, GradNet leads to state-of-the-art performance across different noise levels, while retaining high computational efficiency.

Lacking realistic ground truth data, image denoising techniques are traditionally evaluated on images corrupted by synthesized i.i.d. Gaussian noise. We aim to obviate this unrealistic setting by developing a methodology for benchmarking denoising techniques on real photographs. We capture pairs of images with different ISO values and appropriately adjusted exposure times, where the nearly noise-free low-ISO image serves as reference. To derive the ground truth, careful post-processing is needed. We correct spatial misalignment, cope with inaccuracies in the exposure parameters through a linear intensity transform based on a novel heteroscedastic Tobit regression model, and remove residual low-frequency bias that stems, e.g., from minor illumination changes. We then capture a novel benchmark dataset, the Darmstadt Noise Dataset (DND), with consumer cameras of differing sensor sizes. One interesting finding is that various recent techniques that perform well on synthetic noise are clearly outperformed by BM3D on photographs with real noise. Our benchmark delineates realistic evaluation scenarios that deviate strongly from those commonly used in the scientific literature.

Optical flow estimation is one of the most studied problems in computer vision, yet recent benchmark datasets continue to reveal problem areas of today's approaches. Occlusions have remained one of the key challenges. In this paper, we propose a symmetric optical flow method to address the well-known chicken-and-egg relation between optical flow and occlusions. In contrast to many state-of-the-art methods that consider occlusions as outliers, possibly filtered out during post-processing, we highlight the importance of joint occlusion reasoning in the optimization and show how to utilize occlusion as an important cue for estimating optical flow. The key feature of our model is to fully exploit the symmetry properties that characterize optical flow and occlusions in the two consecutive images. Specifically through utilizing forward-backward consistency and occlusion-disocclusion symmetry in the energy, our model jointly estimates optical flow in both forward and backward direction, as well as consistent occlusion maps in both views. We demonstrate significant performance benefits on standard benchmarks, especially from the occlusion-disocclusion symmetry. On the challenging KITTI dataset we report the most accurate two-frame results to date.

Optical flow estimation remains challenging due to untextured areas, motion boundaries, occlusions, and more. Thus, the estimated flow is not equally reliable across the image. To that end, post-hoc confidence measures have been introduced to assess the per-pixel reliability of the flow. We overcome the artificial separation of optical flow and confidence estimation by introducing a method that jointly predicts optical flow and its underlying uncertainty. Starting from common energy-based formulations, we rely on the corresponding posterior distribution of the flow given the images. We derive a variational inference scheme based on mean field, which incorporates best practices from energy minimization. An uncertainty measure is obtained along the flow at every pixel as the (marginal) entropy of the variational distribution. We demonstrate the flexibility of our probabilistic approach by applying it to two different energies and on two benchmarks. We not only obtain flow results that are competitive with the underlying energy minimization approach, but also a reliable uncertainty measure that significantly outperforms existing post-hoc approaches.

Many existing approaches for image-to-geometry registration assume that either a textured 3D model or a good initial guess of the 3D pose is available to bootstrap the registration process. In this paper we consider the registration of photographs to 3D models even when no texture information is available. This is very challenging as we cannot rely on texture gradients, and even shading gradients are hard to estimate since the lighting conditions are unknown. To that end, we propose average shading gradients , a rendering technique that estimates the average gradient magnitude over all lighting directions under Lambertian shading. We use this gradient representation as the building block of a registration pipeline based on matching sparse features. To cope with inevitable false matches due to the missing texture information and to increase robustness, the pose of the 3D model is estimated in two stages. Coarse pose hypotheses are first obtained from a single correct match each, subsequently refined using SIFT flow, and finally verified. We apply our algorithm to registering images of real-world objects to untextured 3D meshes of limited accuracy. Moreover, we show that registration can be performed even for paintings despite lacking photo-realism.

We propose a novel approach to semantic scene labeling in urban scenarios, which aims to combine excellent recognition performance with highest levels of computational efficiency. To that end, we exploit efficient tree-structured models on two levels: pixels and superpixels. At the pixel level, we propose to unify pixel labeling and the extraction of semantic texton features within a single architecture, so-called encode-and-classify trees. At the superpixel level, we put forward a multi-cue segmentation tree that groups superpixels at multiple granularities. Through learning, the segmentation tree effectively exploits and aggregates a wide range of complementary information present in the data. A tree-structured CRF is then used to jointly infer the labels of all regions across the tree. Finally, we introduce a novel object-centric evaluation method that specifically addresses the urban setting with its strongly varying object scales. Our experiments demonstrate competitive labeling performance compared to the state of the art, while achieving near real-time frame rates of up to 20 fps.

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.

Motion blur can adversely affect a number of vision tasks, hence it is generally considered a nuisance. We instead treat motion blur as a useful signal that allows to compute the motion of objects from a single image. Drawing on the success of joint segmentation and parametric motion models in the context of optical flow estimation, we propose a parametric object motion model combined with a segmentation mask to exploit localized, non-uniform motion blur. Our parametric image formation model is differentiable w.r.t. the motion parameters, which enables us to generalize marginal-likelihood techniques from uniform blind deblurring to localized, non-uniform blur. A two-stage pipeline, first in derivative space and then in image space, allows to estimate both parametric object motion as well as a motion segmentation from a single image alone. Our experiments demonstrate its ability to cope with very challenging cases of object motion blur.

In this paper we present Semantic Stixels, a novel vision-based scene model geared towards automated driving. Our model jointly infers the geometric and semantic layout of a scene and provides a compact yet rich abstraction of both cues using Stixels as primitive elements. Geometric information is incorporated into our model in terms of pixel-level disparity maps derived from stereo vision. For semantics, we leverage a modern deep learning-based scene labeling approach that provides an object class label for each pixel. Our experiments involve an in-depth analysis and a comprehensive assessment of the constituent parts of our approach using three public benchmark datasets. We evaluate the geometric and semantic accuracy of our model and analyze the underlying run-times and the complexity of the obtained representation. Our results indicate that the joint treatment of both cues on the Semantic Stixel level yields a highly compact environment representation while maintaining an accuracy comparable to the two individual pixel-level input data sources. Moreover, our framework compares favorably to related approaches in terms of computational costs and operates in real-time.

The task of tracking multiple targets is often addressed with the so-called tracking-by-detection paradigm, where the first step is to obtain a set of target hypotheses for each frame independently. Tracking can then be regarded as solving two separate, but tightly coupled problems. The first is to carry out data association, i.e., to determine the origin of each of the available observations. The second problem is to reconstruct the actual trajectories that describe the spatio-temporal motion pattern of each individual target. The former is inherently a discrete problem, while the latter should intuitively be modeled in continuous space. Having to deal with an unknown number of targets, complex dependencies, and physical constraints, both are challenging tasks on their own and thus most previous work focuses on one of these subproblems. Here, we present a multi-target tracking approach that explicitly models both tasks as minimization of a unified discrete-continuous energy function. Trajectory properties are captured through global label costs, a recent concept from multi-model fitting, which we introduce to tracking. Specifically, label costs describe physical properties of individual tracks, e.g., linear and angular dynamics, or entry and exit points. We further introduce pairwise label costs to describe mutual interactions between targets in order to avoid collisions. By choosing appropriate forms for the individual energy components, powerful discrete optimization techniques can be leveraged to address data association, while the shapes of individual trajectories are updated by gradient-based continuous energy minimization. The proposed method achieves state-of-the-art results on diverse benchmark sequences.

Conditional random fields (CRFs) are popular discriminative models for computer vision and have been successfully applied in the domain of image restoration, especially to image denoising. For image deblurring, however, discriminative approaches have been mostly lacking. We posit two reasons for this: First, the blur kernel is often only known at test time, requiring any discriminative approach to cope with considerable variability. Second, given this variability it is quite difficult to construct suitable features for discriminative prediction. To address these challenges we first show a connection between common half-quadratic inference for generative image priors and Gaussian CRFs. Based on this analysis, we then propose a cascade model for image restoration that consists of a Gaussian CRF at each stage. Each stage of our cascade is semi-parametric, i.e., it depends on the instance-specific parameters of the restoration problem, such as the blur kernel. We train our model by loss minimization with synthetically generated training data. Our experiments show that when applied to non-blind image deblurring, the proposed approach is efficient and yields state-of-the-art restoration quality on images corrupted with synthetic and real blur. Moreover, we demonstrate its suitability for image denoising, where we achieve competitive results for grayscale and color images.

Estimating surface normals from just a single image is challenging. To simplify the problem, previous work focused on special cases, including directional lighting, known reflectance maps, etc., making shape from shading impractical outside the lab. To cope with more realistic settings, shading cues need to be combined and generalized to natural illumination. This significantly increases the complexity of the approach, as well as the number of parameters that require tuning. Enabled by a new large-scale dataset for training and analysis, we address this with a discriminative learning approach to shape from shading, which uses regression forests for efficient pixel-independent prediction and fast learning. Von Mises-Fisher distributions in the leaves of each tree enable the estimation of surface normals. To account for their expected spatial regularity, we introduce spatial features, including texton and silhouette features. The proposed silhouette features are computed from the occluding contours of the surface and provide scale-invariant context. Aside from computational efficiency, they enable good generalization to unseen data and importantly allow for a robust estimation of the reflectance map, extending our approach to the uncalibrated setting. Experiments show that our discriminative approach outperforms state-of-the-art methods on synthetic and real-world datasets.

Many existing approaches for image-to-geometry registration assume that either a textured 3D model or a good initial guess of the 3D pose is available to bootstrap the registration process. In this paper we consider the registration of photographs to 3D models even when no texture information is available. This is very challenging as we cannot rely on texture gradients, and even shading gradients are hard to estimate since the lighting conditions are unknown. To that end, we propose average shading gradients, a rendering technique that estimates the average gradient magnitude over all lighting directions under Lambertian shading. We use this gradient representation as the building block of a registration pipeline based on matching sparse features. To cope with inevitable false matches due to the missing texture information and to increase robustness, the pose of the 3D model is estimated in two stages. Coarse pose hypotheses are first obtained from a single correct match each, subsequently refined using SIFT flow, and finally verified. We apply our algorithm to registering images of real-world objects to untextured 3D meshes of limited accuracy.

Image blur from camera shake is a common cause for poor image quality in digital photography, prompting a significant recent interest in image deblurring. The vast majority of work on blind deblurring splits the problem into two subsequent steps: First, the blur process (i.e., blur kernel) is estimated, then the image is restored given the estimated kernel using a non-blind deblurring algorithm. Recent work in non-blind deblurring has shown that discriminative approaches can have clear image quality and runtime benefits over typical generative formulations. In this paper, we propose a cascade for blind deblurring that alternates between kernel estimation and discriminative deblurring using regression tree fields (RTFs). We further contribute a new dataset of realistic image blur kernels from human camera shake, which we use to train the discriminative component. Extensive qualitative and quantitative experiments show a clear gain in image quality by interleaving kernel estimation and discriminative deblurring in an iterative cascade.

Estimating surface normals from a single image alone is a challenging problem. Previous work made various simplifications and focused on special cases, such as having directional lighting, known reflectance maps, etc. This is problematic, however, as shape from shading becomes impractical outside the lab. We argue that addressing more realistic settings requires multiple shading cues to be combined as well as generalized to natural illumination. However, this requires coping with an increased complexity of the approach and more parameters to be adjusted. Starting from a novel large-scale dataset for training and analysis, we pursue a discriminative learning approach to shape from shading. Regression forests enable efficient pixel-independent prediction and fast learning. The regression trees are adapted to predicting surface normals by using von Mises–Fisher distributions in the leaves. Spatial regularity of the normals is achieved through a combination of spatial features, including texton as well as novel silhouette features. The proposed silhouette features leverage the occluding contours of the surface and yield scale-invariant context. Their benefits include computational efficiency and good generalization to unseen data. Importantly, they allow estimating the reflectance map robustly, thus addressing the uncalibrated setting. Our method can also be extended to handle perspective projection. Experiments show that our discriminative approach outperforms the state of the art on various synthetic and real-world datasets.

Many state-of-the-art image restoration approaches do not scale well to larger images, such as megapixel images common in the consumer segment. Computationally expensive optimization is often the culprit. While efficient alternatives exist, they have not reached the same level of image quality. The goal of this paper is to develop an effective approach to image restoration that offers both computational efficiency and high restoration quality. To that end we propose shrinkage fields, a random field-based architecture that combines the image model and the optimization algorithm in a single unit. The underlying shrinkage operation bears connections to wavelet approaches, but is used here in a random field context. Computational efficiency is achieved by construction through the use of convolution and DFT as the core components, high restoration quality is attained through loss-based training of all model parameters and the use of a cascade architecture. Unlike heavily engineered solutions, our learning approach can be adapted easily to different trade-offs between efficiency and image quality. We demonstrate state-of-the-art restoration results with high levels of computational efficiency, and significant speedup potential through inherent parallelism.

We address the problem of estimating and removing localized image blur, as it for example arises from moving objects in a scene, or when the depth of field is insufficient to sharply render all objects of interest. Unlike the case of camera shake, such blur changes abruptly at the object boundaries. To cope with this, we propose an automated sharp image recovery method that simultaneously determines blurred regions and estimates their responsible blur kernels. To address a wide range of different scenarios, our model is not restricted to a discrete set of candidate blurs, but allows for arbitrary, non-parametric blur kernels. Moreover, our approach does not require specialized hardware, an alpha matte, or user annotation of the blurred region. Unlike previous methods, we show that localized blur estimation can be accomplished by incorporating a pixel-wise latent variable to indicate the active blur kernel. Furthermore, we generalize the marginal likelihood technique of blind deblurring to the case of localized blur. Specifically, we integrate out the latent image derivatives to permit marginal density estimates of both blur kernels and their regions of influence. We obtain sharp images in applications to both object motion blur and defocus blur removal. Quantitative results on two novel datasets as well as qualitative results comparing to a range of specialized methods demonstrate the versatility and effectiveness of our non-parametric approach.

The accuracy of optical flow estimation algorithms has been improving steadily as evidenced by results on the Middlebury optical flow benchmark. The typical formulation, however, has changed little since the work of Horn and Schunck. We attempt to uncover what has made recent advances possible through a thorough analysis of how the objective function, the optimization method, and modern implementation practices influence accuracy. We discover that “classical” flow formulations perform surprisingly well when combined with modern optimization and implementation techniques. One key implementation detail is the median filtering of intermediate flow fields during optimization. While this improves the robustness of classical methods it actually leads to higher energy solutions, meaning that these methods are not optimizing the original objective function. To understand the principles behind this phenomenon, we derive a new objective function that formalizes the median filtering heuristic. This objective function includes a non-local smoothness term that robustly integrates flow estimates over large spatial neighborhoods. By modifying this new term to include information about flow and image boundaries we develop a method that can better preserve motion details. To take advantage of the trend towards video in wide-screen format, we further introduce an asymmetric pyramid downsampling scheme that enables the estimation of longer range horizontal motions. The methods are evaluated on the Middlebury, MPI Sintel, and KITTI datasets using the same parameter settings.

Many recent advances in multiple target tracking aim at finding a (nearly) optimal set of trajectories within a temporal window. To handle the large space of possible trajectory hypotheses, it is typically reduced to a finite set by some form of data-driven or regular discretization. In this work, we propose an alternative formulation of multitarget tracking as minimization of a continuous energy. Contrary to recent approaches, we focus on designing an energy that corresponds to a more complete representation of the problem, rather than one that is amenable to global optimization. Besides the image evidence, the energy function takes into account physical constraints, such as target dynamics, mutual exclusion, and track persistence. In addition, partial image evidence is handled with explicit occlusion reasoning, and different targets are disambiguated with an appearance model. To nevertheless find strong local minima of the proposed nonconvex energy, we construct a suitable optimization scheme that alternates between continuous conjugate gradient descent and discrete transdimensional jump moves. These moves, which are executed such that they always reduce the energy, allow the search to escape weak minima and explore a much larger portion of the search space of varying dimensionality. We demonstrate the validity of our approach with an extensive quantitative evaluation on several public data sets.

When tracking multiple targets in crowded scenarios, modeling mutual exclusion between distinct targets becomes important at two levels: (1) in data association, each target observation should support at most one trajectory and each trajectory should be assigned at most one observation per frame, (2) in trajectory estimation, two trajectories should remain spatially separated at all times to avoid collisions. Yet, existing trackers often sidestep these important constraints. We address this using a mixed discrete-continuous conditional random field (CRF) that explicitly models both types of constraints: Exclusion between conflicting observations with super modular pairwise terms, and exclusion between trajectories by generalizing global label costs to suppress the co-occurrence of incompatible labels (trajectories). We develop an expansion move-based MAP estimation scheme that handles both non-sub modular constraints and pairwise global label costs. Furthermore, we perform a statistical analysis of ground-truth trajectories to derive appropriate CRF potentials for modeling data fidelity, target dynamics, and inter-target occlusion.

Non-blind deblurring is an integral component of blind approaches for removing image blur due to camera shake. Even though learning-based deblurring methods exist, they have been limited to the generative case and are computationally expensive. To this date, manually-defined models are thus most widely used, though limiting the attained restoration quality. We address this gap by proposing a discriminative approach for non-blind deblurring. One key challenge is that the blur kernel in use at test time is not known in advance. To address this, we analyze existing approaches that use half-quadratic regularization. From this analysis, we derive a discriminative model cascade for image deblurring. Our cascade model consists of a Gaussian CRF at each stage, based on the recently introduced regression tree fields. We train our model by loss minimization and use synthetically generated blur kernels to generate training data. Our experiments show that the proposed approach is efficient and yields state-of-the-art restoration quality on images corrupted with synthetic and real blur.

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.

Estimating dense 3D scene flow from stereo sequences remains a challenging task, despite much progress in both classical disparity and 2D optical flow estimation. To overcome the limitations of existing techniques, we introduce a novel model that represents the dynamic 3D scene by a collection of planar, rigidly moving, local segments. Scene flow estimation then amounts to jointly estimating the pixel-to-segment assignment, and the 3D position, normal vector, and rigid motion parameters of a plane for each segment. The proposed energy combines an occlusion-sensitive data term with appropriate shape, motion, and segmentation regularizers. Optimization proceeds in two stages: Starting from an initial super pixelization, we estimate the shape and motion parameters of all segments by assigning a proposal from a set of moving planes. Then the pixel-to-segment assignment is updated, while holding the shape and motion parameters of the moving planes fixed. We demonstrate the benefits of our model on different real-world image sets, including the challenging KITTI benchmark. We achieve leading performance levels, exceeding competing 3D scene flow methods, and even yielding better 2D motion estimates than all tested dedicated optical flow techniques.

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.

The problem of multi-target tracking is comprised of two distinct, but tightly coupled challenges: (i) the naturally discrete problem of data association, i.e. assigning image observations to the appropriate target; (ii) the naturally continuous problem of trajectory estimation, i.e. recovering the trajectories of all targets. To go beyond simple greedy solutions for data association, recent approaches often perform multi-target tracking using discrete optimization. This has the disadvantage that trajectories need to be pre-computed or represented discretely, thus limiting accuracy. In this paper we instead formulate multi-target tracking as a discrete-continuous optimization problem that handles each aspect in its natural domain and allows leveraging powerful methods for multi-model fitting. Data association is performed using discrete optimization with label costs, yielding near optimality. Trajectory estimation is posed as a continuous fitting problem with a simple closed-form solution, which is used in turn to update the label costs. We demonstrate the accuracy and robustness of our approach with state-of-the-art performance on several standard datasets.

Identifying suitable image features is a central challenge in computer vision, ranging from representations for low-level to high-level vision. Due to the difficulty of this task, techniques for learning features directly from example data have recently gained attention. Despite significant benefits, these learned features often have many fewer of the desired invariances or equivariances than their hand-crafted counterparts. While translation in-/equivariance has been addressed, the issue of learning rotation-invariant or equivariant representations is hardly explored. In this paper we describe a general framework for incorporating invariance to linear image transformations into product models for feature learning. A particular benefit is that our approach induces transformation-aware feature learning, i.e. it yields features that have a notion with which specific image transformation they are used. We focus our study on rotation in-/equivariance and show the advantages of our approach in learning rotation-invariant image priors and in building rotation-equivariant and invariant descriptors of learned features, which result in state-of-the-art performance for rotation-invariant object detection.

Spatially-discrete Markov random fields (MRFs) and spatially-continuous variational approaches are ubiquitous in low-level vision, including image restoration, segmentation, optical flow, and stereo. Even though both families of approaches are fairly similar on an intuitive level, they are frequently seen as being technically rather distinct since they operate on different domains. In this paper we explore their connections and develop a direct, rigorous link with a particular emphasis on first-order regularizers. By representing spatially-continuous functions as linear combinations of finite elements with local support and performing explicit integration of the variational objective, we derive MRF potentials that make the resulting MRF energy equivalent to the variational energy functional. In contrast to previous attempts, we provide an explicit connection for modern non-quadratic regularizers and also integrate the data term. The established connection opens certain classes of MRFs to spatially-continuous interpretations and variational formulations to a broad range of probabilistic learning and inference algorithms.

Conventional non-blind image deblurring algorithms involve natural image priors and maximum a-posteriori (MAP) estimation. As a consequence of MAP estimation, separate pre-processing steps such as noise estimation and training of the regularization parameter are necessary to avoid user interaction. Moreover, MAP estimates involving standard natural image priors have been found lacking in terms of restoration performance. To address these issues we introduce an integrated Bayesian framework that unifies non-blind deblurring and noise estimation, thus freeing the user of tediously pre-determining a noise level. A sampling-based technique allows to integrate out the unknown noise level and to perform deblurring using the Bayesian minimum mean squared error estimate (MMSE), which requires no regularization parameter and yields higher performance than MAP estimates when combined with a learned high-order image prior. A quantitative evaluation demonstrates state-of-the-art results for both non-blind deblurring and noise estimation.

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.

We present an approach to 3D scene flow estimation, which exploits that in realistic scenarios image motion is frequently dominated by observer motion and independent, but rigid object motion. We cast the dense estimation of both scene structure and 3D motion from sequences of two or more views as a single energy minimization problem. We show that agnostic smoothness priors, such as the popular total variation, are biased against motion discontinuities in viewing direction. Instead, we propose to regularize by encouraging local rigidity of the 3D scene. We derive a local rigidity constraint of the 3D scene flow and define a smoothness term that penalizes deviations from that constraint, thus favoring solutions that consist largely of rigidly moving parts. Our experiments show that the new rigid motion prior reduces the 3D flow error by 42% compared to standard TV regularization with the same data term.

The quantitative evaluation of optical flow algorithms by Barron et al. (1994) led to significant advances in performance. The challenges for optical flow algorithms today go beyond the datasets and evaluation methods proposed in that paper. Instead, they center on problems associated with complex natural scenes, including nonrigid motion, real sensor noise, and motion discontinuities. We propose a new set of benchmarks and evaluation methods for the next generation of optical flow algorithms. To that end, we contribute four types of data to test different aspects of optical flow algorithms: (1) sequences with nonrigid motion where the ground-truth flow is determined by tracking hidden fluorescent texture, (2) realistic synthetic sequences, (3) high frame-rate video used to study interpolation error, and (4) modified stereo sequences of static scenes. In addition to the average angular error used by Barron et al., we compute the absolute flow endpoint error, measures for frame interpolation error, improved statistics, and results at motion discontinuities and in textureless regions. In October 2007, we published the performance of several well-known methods on a preliminary version of our data to establish the current state of the art. We also made the data freely available on the web at http://vision.middlebury.edu/flow/ . Subsequently a number of researchers have uploaded their results to our website and published papers using the data. A significant improvement in performance has already been achieved. In this paper we analyze the results obtained to date and draw a large number of conclusions from them.

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.

Markov random fields (MRFs) are popular and generic probabilistic models of prior knowledge in low-level vision. Yet their generative properties are rarely examined, while application-specific models and non-probabilistic learning are gaining increased attention. In this paper we revisit the generative aspects of MRFs, and analyze the quality of common image priors in a fully application-neutral setting. Enabled by a general class of MRFs with flexible potentials and an efficient Gibbs sampler, we find that common models do not capture the statistics of natural images well. We show how to remedy this by exploiting the efficient sampler for learning better generative MRFs based on flexible potentials. We perform image restoration with these models by computing the Bayesian minimum mean squared error estimate (MMSE) using sampling. This addresses a number of shortcomings that have limited generative MRFs so far, and leads to substantially improved performance over maximum a-posteriori (MAP) estimation. We demonstrate that combining our learned generative models with sampling-based MMSE estimation yields excellent application results that can compete with recent discriminative methods.

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.

The accuracy of optical flow estimation algorithms has been improving steadily as evidenced by results on the Middlebury optical flow benchmark. The typical formulation, however, has changed little since the work of Horn and Schunck. We attempt to uncover what has made recent advances possible through a thorough analysis of how the objective function, the optimization method, and modern implementation practices influence accuracy. We discover that “classical” flow formulations perform surprisingly well when combined with modern optimization and implementation techniques. Moreover, we find that while median filtering of intermediate flow fields during optimization is a key to recent performance gains, it leads to higher energy solutions. To understand the principles behind this phenomenon, we derive a new objective that formalizes the median filtering heuristic. This objective includes a nonlocal term that robustly integrates flow estimates over large spatial neighborhoods. By modifying this new term to include information about flow and image boundaries we develop a method that ranks at the top of the Middlebury benchmark.

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.

The efficient application of graph cuts to Markov Random Fields (MRFs) with multiple discrete or continuous labels remains an open question. In this paper, we demonstrate one possible way of achieving this by using graph cuts to combine pairs of suboptimal labelings or solutions. We call this combination process the fusion move. By employing recently developed graph-cut-based algorithms (so-called QPBO-graph cut), the fusion move can efficiently combine two proposal labelings in a theoretically sound way, which is in practice often globally optimal. We demonstrate that fusion moves generalize many previous graph-cut approaches, which allows them to be used as building blocks within a broader variety of optimization schemes than were considered before. In particular, we propose new optimization schemes for computer vision MRFs with applications to image restoration, stereo, and optical flow, among others. Within these schemes the fusion moves are used 1) for the parallelization of MRF optimization into several threads, 2) for fast MRF optimization by combining cheap-to-compute solutions, and 3) for the optimization of highly nonconvex continuous-labeled MRFs with 2D labels. Our final example is a nonvision MRF concerned with cartographic label placement, where fusion moves can be used to improve the performance of a standard inference method (loopy belief propagation).

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.

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.

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.

Accurate estimation of optical flow is a challenging task, which often requires addressing difficult energy optimization problems. To solve them, most top-performing methods rely on continuous optimization algorithms. The modeling accuracy of the energy in this case is often traded for its tractability. This is in contrast to the related problem of narrow-baseline stereo matching, where the top-performing methods employ powerful discrete optimization algorithms such as graph cuts and message-passing to optimize highly non-convex energies. In this paper, we demonstrate how similar non-convex energies can be formulated and optimized discretely in the context of optical flow estimation. Starting with a set of candidate solutions that are produced by fast continuous flow estimation algorithms, the proposed method iteratively fuses these candidate solutions by the computation of minimum cuts on graphs. The obtained continuous-valued fusion result is then further improved using local gradient descent. Experimentally, we demonstrate that the proposed energy is an accurate model and that the proposed discrete-continuous optimization scheme not only finds lower energy solutions than traditional discrete or continuous optimization techniques, but also leads to flow estimates that outperform the current state-of-the-art.

The quantitative evaluation of optical flow algorithms by Barron et al. led to significant advances in the performance of optical flow methods. The challenges for optical flow today go beyond the datasets and evaluation methods proposed in that paper and center on problems associated with nonrigid motion, real sensor noise, complex natural scenes, and motion discontinuities. Our goal is to establish a new set of benchmarks and evaluation methods for the next generation of optical flow algorithms. To that end, we contribute four types of data to test different aspects of optical flow algorithms: sequences with nonrigid motion where the ground-truth flow is determined by tracking hidden fluorescent texture; realistic synthetic sequences; high frame-rate video used to study interpolation error; and modified stereo sequences of static scenes. In addition to the average angular error used in Barron et al., we compute the absolute flow endpoint error, measures for frame interpolation error, improved statistics, and flow accuracy at motion boundaries and in textureless regions. We evaluate the performance of several well-known methods on this data to establish the current state of the art. Our database is freely available on the web together with scripts for scoring and publication of the results at http://vision.middlebury.edu/flow/.

In contrast to traditional Markov random field (MRF) models, we develop a steerable random field (SRF) in which the field potentials are defined in terms of filter responses that are steered to the local image structure. In particular, we use the structure tensor to obtain derivative responses that are either aligned with, or orthogonal to, the predominant local image structure, and analyze the statistics of these steered filter responses in natural images. Clique potentials are defined over steered filter responses using a Gaussian scale mixture model and are learned from training data. The SRF model connects random field models with anisotropic regularization and provides a statistical motivation for the latter. We demonstrate that steering the random field to the local image structure improves image denoising and inpainting performance compared with traditional pairwise MRFs.

In scenes containing specular objects, the image motion observed by a moving camera may be an intermixed combination of optical flow resulting from diffuse reflectance (diffuse flow) and specular reflection (specular flow). Here, with few assumptions, we formalize the notion of specular flow, show how it relates to the 3D structure of the world, and develop an algorithm for estimating scene structure from 2D image motion. Unlike previous work on isolated specular highlights we use two image frames and estimate the semi-dense flow arising from the specular reflections of textured scenes. We parametrically model the image motion of a quadratic surface patch viewed from a moving camera. The flow is modeled as a probabilistic mixture of diffuse and specular components and the 3D shape is recovered using an Expectation-Maximization algorithm. Rather than treating specular reflections as noise to be removed or ignored, we show that the specular flow provides additional constraints on scene geometry that improve estimation of 3D structure when compared with reconstruction from diffuse flow alone. We demonstrate this for a set of synthetic and real sequences of mixed specular-diffuse objects.

We develop a Bayesian model of digitized archival films and use this for denoising, or more specifically de-graining, individual frames. In contrast to previous approaches our model uses a learned spatial prior and a unique likelihood term that models the physics that generates the image grain. The spatial prior is represented by a high-order Markov random field based on the recently proposed field-of-experts framework. We propose a new model of the image grain in archival films based on an inhomogeneous beta distribution in which the variance is a function of image luminance. We train this noise model for a particular film and perform de-graining using a diffusion method. Quantitative results show improved signal-to-noise ratio relative to the standard ad hoc Gaussian noise model.

We develop a framework for learning generic, expressive image priors that capture the statistics of natural scenes and can be used for a variety of machine vision tasks. The approach extends traditional Markov random field (MRF) models by learning potential functions over extended pixel neighborhoods. Field potentials are modeled using a Products-of-Experts framework that exploits nonlinear functions of many linear filter responses. In contrast to previous MRF approaches all parameters, including the linear filters themselves, are learned from training data. We demonstrate the capabilities of this Field of Experts model with two example applications, image denoising and image inpainting, which are implemented using a simple, approximate inference scheme. While the model is trained on a generic image database and is not tuned toward a specific application, we obtain results that compete with and even outperform specialized techniques.

We develop a method for learning the spatial statistics of optical flow fields from a novel training database. Training flow fields are constructed using range images of natural scenes and 3D camera motions recovered from handheld and car-mounted video sequences. A detailed analysis of optical flow statistics in natural scenes is presented and machine learning methods are developed to learn a Markov random field model of optical flow. The prior probability of a flow field is formulated as a field-of-experts model that captures the higher order spatial statistics in overlapping patches and is trained using contrastive divergence. This new optical flow prior is compared with previous robust priors and is incorporated into a recent, accurate algorithm for dense optical flow computation. Experiments with natural and synthetic sequences illustrate how the learned optical flow prior quantitatively improves flow accuracy and how it captures the rich spatial structure found in natural scene motion.

Probabilistic modeling of correlated neural population firing activity is central to understanding the neural code and building practical decoding algorithms. No parametric models currently exist for modeling multivariate correlated neural data and the high dimensional nature of the data makes fully non-parametric methods impractical. To address these problems we propose an energy-based model in which the joint probability of neural activity is represented using learned functions of the 1D marginal histograms of the data. The parameters of the model are learned using contrastive divergence and an optimization procedure for finding appropriate marginal directions. We evaluate the method using real data recorded from a population of motor cortical neurons. In particular, we model the joint probability of population spiking times and 2D hand position and show that the likelihood of test data under our model is significantly higher than under other models. These results suggest that our model captures correlations in the firing activity. Our rich probabilistic model of neural population activity is a step towards both measurement of the importance of correlations in neural coding and improved decoding of population activity.