Out-of-distribution (OOD) detection is crucial for deploying robust and reliable machine-learning systems in open-world settings. Despite steady advances in OOD detectors, their interplay with modern training pipelines that maximize in-distribution (ID) accuracy and generalization remains under-explored. We investigate this link through a comprehensive empirical study. Fixing the architecture to the widely adopted ResNet-50, we benchmark 21 post-hoc, state-of-the-art OOD detection methods across 56 ImageNet-trained models obtained via diverse training strategies and evaluate them on eight OOD test sets. Contrary to the common assumption that higher ID accuracy implies better OOD detection performance, we uncover a non-monotonic relationship: OOD performance initially improves with accuracy but declines once advanced training recipes push accuracy beyond the baseline. Moreover, we observe a strong interdependence between training strategy, detector choice, and resulting OOD performance, indicating that no single method is universally optimal.

Current progress in out-of-distribution (OOD) detection is limited by the lack of large, high-quality datasets with clearly defined OOD categories across varying difficulty levels (near- to far-OOD) that support both fine- and coarse-grained computer vision tasks. To address this limitation, we introduce ICONIC-444 (Image Classification and OOD Detection with Numerous Intricate Complexities), a specialized large-scale industrial image dataset containing over 3.1 million RGB images spanning 444 classes tailored for OOD detection research. Captured with a prototype industrial sorting machine, ICONIC-444 closely mimics real-world tasks. It complements existing datasets by offering structured, diverse data suited for rigorous OOD evaluation across a spectrum of task complexities. We define four reference tasks within ICONIC-444 to benchmark and advance OOD detection research and provide baseline results for 22 state-of-the-art post-hoc OOD detection methods.

Future trajectories of neighboring traffic agents have a significant influence on the path planning and decision-making of autonomous vehicles. While trajectory forecasting is a well-studied field, research mainly focuses on snapshot-based prediction, where each scenario is treated independently of its global temporal context. However, real-world autonomous driving systems need to operate in a continuous setting, requiring real-time processing of data streams with low latency and consistent predictions over successive timesteps. We leverage this continuous setting to propose a lightweight yet highly accurate streaming-based trajectory forecasting approach. We integrate valuable information from previous predictions with a novel endpoint-aware modeling scheme. Our temporal context propagation uses the trajectory endpoints of the previous forecasts as anchors to extract targeted scenario context encodings. Our approach efficiently guides its scene encoder to extract highly relevant context information without needing refinement iterations or segment-wise decoding. Our experiments highlight that our approach effectively relays information across consecutive timesteps. Unlike methods using multi-stage refinement processing, our approach significantly reduces inference latency, making it well-suited for real-world deployment. We achieve state-of-the-art streaming trajectory prediction results on the Argoverse 2 multi-agent and single-agent benchmarks, while requiring substantially fewer resources.

While surface normals are widely used to analyse 3D scene geometry, surface normal estimation from LiDAR point clouds remains severely underexplored. This is caused by the lack of large-scale annotated datasets on the one hand, and lack of methods that can robustly handle the sparse and often noisy LiDAR data in a reasonable time on the other hand. We address these limitations using a traffic simulation engine and present LiSu, the first large-scale, synthetic LiDAR point cloud dataset with ground truth surface normal annotations, eliminating the need for tedious manual labeling. Additionally, we propose a novel method that exploits the spatiotemporal characteristics of autonomous driving data to enhance surface normal estimation accuracy. By incorporating two regularization terms, we enforce spatial consistency among neighboring points and temporal smoothness across consecutive LiDAR frames. These regularizers are particularly effective in self-training settings, where they mitigate the impact of noisy pseudo-labels, enabling robust real-world deployment. We demonstrate the effectiveness of our method on LiSu, achieving state-of-the-art performance in LiDAR surface normal estimation. Moreover, we showcase its full potential in addressing the challenging task of synthetic-to-real domain adaptation, leading to improved neural surface reconstruction on real-world data.

LiDAR-based 3D detectors need large datasets for training, yet they struggle to generalize to novel domains. Domain Generalization (DG) aims to mitigate this by training detectors that are invariant to such domain shifts. Current DG approaches exclusively rely on global geometric features (point cloud Cartesian coordinates) as input features. Over-reliance on these global geometric features can, however, cause 3D detectors to prioritize object location and absolute position, resulting in poor cross-domain performance. To mitigate this, we propose to exploit explicit local point cloud structure for DG, in particular by encoding point cloud neighborhoods with Gaussian blobs, GBlobs. Our proposed formulation is highly efficient and requires no additional parameters. Without any bells and whistles, simply by integrating GBlobs in existing detectors, we beat the current state-of-the-art in challenging single-source DG benchmarks by over 21 mAP (Waymo→KITTI), 13 mAP (KITTI→Waymo), and 12 mAP (nuScenes→KITTI), without sacrificing in-domain performance. Additionally, GBlobs demonstrate exceptional performance in multi-source DG, surpassing the current state-of-the-art by 17, 12, and 5 mAP on Waymo, KITTI, and ONCE, respectively.

Trajectory prediction is an essential component of the perception stack in autonomous mobile robots (AMRs). AMRs operate in complex environments where their movements are influenced by various environment elements, such as racks and storage locations. Therefore, accurate and efficient trajectory prediction for intralogistics requires detailed environment modeling that goes beyond the lane-based context mainly used in road traffic methods. We propose the addition of a new environment context encoder module that can be seamlessly integrated into state-of-the-art autonomous driving systems. Our approach, tailored to the specific challenges of intralogistics, achieves highly accurate predictions using compact and efficient baseline networks.

Trajectory prediction for autonomous driving has been extensively studied using large-scale datasets from the US and Asia. These datasets typically have a strong bias toward predicting vehicle motion. Recently, the View-of-Delft Prediction (VoD-P) dataset introduced a collection of European urban mixed-traffic scenarios, posing unique challenges due to its diversity and relatively small dataset size. In this work, we conduct a detailed study on trajectory prediction on the VoD-P dataset. We show that state-of-the-art trajectory prediction models, which perform well on large-scale vehicle-biased datasets, struggle to generalize to the scenarios. To address this limitation, we propose a simple yet effective transformer-based trajectory prediction model, specifically designed to handle the challenges posed in diverse urban scenarios. Combining a strong baseline with refined environment modeling, our approach significantly outperforms all existing methods on the VoD-P dataset.

We introduce STSBench, a scenario-based framework to benchmark the holistic understanding of vision-language models (VLMs) for autonomous driving. The framework automatically mines predefined traffic scenarios from any dataset using ground-truth annotations, provides an intuitive user interface for efficient human verification, and generates multiple-choice questions for model evaluation. Applied to the nuScenes dataset, we present STSnu, the first benchmark that evaluates the spatio-temporal reasoning capabilities of VLMs based on comprehensive 3D perception. Existing benchmarks typically target off-the-shelf or fine-tuned VLMs for images or videos from a single viewpoint, focusing on semantic tasks such as object recognition, dense captioning, risk assessment, or scene understanding. In contrast, STSnu evaluates driving expert VLMs for end-to-end driving, operating on videos from multi-view cameras or LiDAR. It specifically assesses their ability to reason about both ego-vehicle actions and complex interactions among traffic participants, a crucial capability for autonomous vehicles. The benchmark features 43 diverse scenarios spanning multiple views and frames, resulting in 971 human-verified multiple-choice questions. A thorough evaluation uncovers critical shortcomings in existing models’ ability to reason about fundamental traffic dynamics in complex environments. These findings highlight the urgent need for architectural advancements that explicitly model spatio-temporal reasoning. By addressing a core gap in spatio-temporal evaluation, STSBench enables the development of more robust and explainable VLMs for autonomous driving.

We propose a novel approach to video anomaly detection: we treat feature vectors extracted from videos as re-alizations of a random variable with a fixed distribution and model this distribution with a neural network. This lets us estimate the likelihood of test videos and detect video anomalies by thresholding the likelihood estimates. We train our video anomaly detector using a modification of de-noising score matching, a method that injects training data with noise to facilitate modeling its distribution. To elim-inate hyperparameter selection, we model the distribution of noisy video features across a range of noise levels and introduce a regularizer that tends to align the models for different levels of noise. At test time, we combine anomaly indications at multiple noise scales with a Gaussian mix-ture model. Running our video anomaly detector induces minimal delays as inference requires merely extracting the features and forward-propagating them through a shallow neural network and a Gaussian mixture model. Our ex-periments on five popular video anomaly detection bench-marks demonstrate state-of-the-art performance, both in the object-centric and in the frame-centric setup.

Using neural networks for localization of key fob within and surrounding a car as a security feature for keyless entry is fast emerging. In this paper we study: 1) the performance of pre-computed features of neural networks based UWB (ultra wide band) localization classification forming the baseline of our experiments. 2) Investigate the inherent robustness of various neural networks; therefore, we include the study of robustness of the adversarial examples without any adversarial training in this work. 3) Propose a multi-head self-supervised neural network architecture which outperforms the baseline neural networks without any adversarial training. The model’s performance improved by 67% at certain ranges of adversarial magnitude for fast gradient sign method and 37% each for basic iterative method and projected gradient descent method.

Understanding actions of other agents increases the efficiency of autonomous mobile robots (AMRs) since they encompass intention and indicate future movements. We propose a new method that allows us to infer vehicle actions using a shallow image-based classification model. The actions are classified via bird’s-eye view scene crops, where we project the detections of a 3D object detection model onto a context map. We learn map context information and aggregate temporal sequence information without requiring object tracking. This results in a highly efficient classification model that can easily be deployed on embedded AMR hardware. To evaluate our approach, we create new large-scale synthetic datasets showing warehouse traffic based on real vehicle models and geometry.

For efficient and safe autonomous driving, it is essential that autonomous vehicles can predict the motion of other traffic agents. While highly accurate, current motion prediction models often impose significant challenges in terms of training resource requirements and deployment on embedded hardware. We propose a new efficient motion prediction model, which achieves highly competitive benchmark results while training only a few hours on a single GPU. Due to our lightweight architectural choices and the focus on reducing the required training resources, our model can easily be applied to custom datasets. Furthermore, its low inference latency makes it particularly suitable for deployment in autonomous applications with limited computing resources.

The sensing process of large-scale LiDAR point clouds inevitably causes large blind spots, i.e. regions not visible to the sensor. We demonstrate how these inherent sampling properties can be effectively utilized for self-supervised representation learning by designing a highly effective pretraining framework that considerably reduces the need for tedious 3D annotations to train state-of-the-art object detectors. Our Masked AutoEncoder for LiDAR point clouds (MAELi) intuitively leverages the sparsity of LiDAR point clouds in both the encoder and decoder during reconstruction. This results in more expressive and useful initialization, which can be directly applied to downstream perception tasks, such as 3D object detection or semantic segmentation for autonomous driving. In a novel reconstruction approach, MAELi distinguishes between empty and occluded space and employs a new masking strategy that targets the LiDAR’s inherent spherical projection. Thereby, without any ground truth whatsoever and trained on single frames only, MAELi obtains an understanding of the underlying 3D scene geometry and semantics. To demonstrate the potential of MAELi, we pre-train backbones in an end-to-end manner and show the effectiveness of our unsupervised pre-trained weights on the tasks of 3D object detection and semantic segmentation.

Out-of-distribution (OOD) detection is essential to ensure the reliability and robustness of machine learning models in real-world applications. Post-hoc OOD detection methods have gained significant attention due to the fact that they offer the advantage of not requiring additional re-training, which could degrade model performance and increase training time. However, most existing post-hoc methods rely only on the encoder output (features), logits, or the softmax probability, meaning they have no access to information that might be lost in the feature extraction process. In this work, we address this limitation by introducing Adaptive Temperature Scaling (ATS), a novel approach that dynamically calculates a temperature value based on activations of the intermediate layers. Fusing this sample-specific adjustment with class-dependent logits, our ATS captures additional statistical information before they are lost in the feature extraction process, leading to a more robust and powerful OOD detection method. We conduct extensive experiments to demonstrate the efficacy of our approach. Notably, our method can be seamlessly combined with SOTA post-hoc OOD detection methods that rely on the logits, thereby enhancing their performance and improving their robustness.

Although action recognition systems can achieve top performance when evaluated on in-distribution test points, they are vulnerable to unanticipated distribution shifts in test data. However, test-time adaptation of video action recognition models against common distribution shifts has so far not been demonstrated. We propose to address this problem with an approach tailored to spatio-temporal models that is capable of adaptation on a single video sample at a step. It consists in a feature distribution alignment technique that aligns online estimates of test set statistics towards the training statistics. We further enforce prediction consistency over temporally augmented views of the same test video sample. Evaluations on three benchmark action recognition datasets show that our proposed technique is architecture-agnostic and able to significantly boost the performance on both, the state of the art convolutional architecture TANet and the Video Swin Transformer. Our proposed method demonstrates a substantial performance gain over existing test-time adaptation approaches in both evaluations of a single distribution shift and the challenging case of random distribution shifts. Code will be available at https://github.com/wlin-at/ViTTA.

Test-Time-Training (TTT) is an approach to cope with out-of-distribution (OOD) data by adapting a trained model to distribution shifts occurring at test-time. We propose to perform this adaptation via Activation Matching (ActMAD): We analyze activations of the model and align activation statistics of the OOD test data to those of the training data. In contrast to existing methods, which model the distribution of entire channels in the ultimate layer of the feature extractor, we model the distribution of each feature in multiple layers across the network. This results in a more fine-grained supervision and makes ActMAD attain state of the art performance on CIFAR-100C and Imagenet-C. ActMAD is also architecture-and task-agnostic, which lets us go beyond image classification, and score 15.4% improvement over previous approaches when evaluating a KITTI-trained object detector on KITTI-Fog. Our experiments highlight that ActMAD can be applied to online adaptation in realistic scenarios, requiring little data to attain its full performance.

Large scale Vision Language (VL) models have shown tremendous success in aligning representations between visual and text modalities. This enables remarkable progress in zero-shot recognition, image generation & editing, and many other exciting tasks. However, VL models tend to over-represent objects while paying much less attention to verbs, and require additional tuning on video data for best zero-shot action recognition performance. While previous work relied on large-scale, fully-annotated data, in this work we propose an unsupervised approach. We adapt a VL model for zero-shot and few-shot action recognition using a collection of unlabeled videos and an unpaired action dictionary. Based on that, we leverage Large Language Models and VL models to build a text bag for each unlabeled video via matching, text expansion and captioning. We use those bags in a Multiple Instance Learning setup to adapt an image-text backbone to video data. Although finetuned on unlabeled video data, our resulting models demonstrate high transferability to numerous unseen zero-shot downstream tasks, improving the base VL model performance by up to 14%, and even comparing favorably to fully-supervised baselines in both zero-shot and few-shot video recognition transfer. The code is released at https://github.com/wlin-at/MAXI.

Our MATE is the first Test-Time-Training (TTT) method designed for 3D data, which makes deep networks trained for point cloud classification robust to distribution shifts occurring in test data. Like existing TTT methods from the 2D image domain, MATE also leverages test data for adaptation. Its test-time objective is that of a Masked Autoencoder: a large portion of each test point cloud is removed before it is fed to the network, tasked with reconstructing the full point cloud. Once the network is updated, it is used to classify the point cloud. We test MATE on several 3D object classification datasets and show that it significantly improves robustness of deep networks to several types of corruptions commonly occurring in 3D point clouds. We show that MATE is very efficient in terms of the fraction of points it needs for the adaptation. It can effectively adapt given as few as 5% of tokens of each test sample, making it extremely lightweight. Our experiments show that MATE also achieves competitive performance by adapting sparsely on the test data, which further reduces its computational overhead, making it ideal for real-time applications.

Widely-used LiDAR-based 3D object detectors often neglect fundamental geometric information readily available from the object proposals in their confidence estimation. This is mostly due to architectural design choices, which were often adopted from the 2D image domain, where geometric context is rarely available. In 3D, however, considering the object properties and its surroundings in a holistic way is important to distinguish between true and false positive detections, e.g. occluded pedestrians in a group. To address this, we present GACE, an intuitive and highly efficient method to improve the confidence estimation of a given black-box 3D object detector. We aggregate geometric cues of detections and their spatial relationships, which enables us to properly assess their plausibility and consequently, improve the confidence estimation. This leads to consistent performance gains over a variety of state-of-the-art detectors. Across all evaluated detectors, GACE proves to be especially beneficial for the vulnerable road user classes, i.e. pedestrians and cyclists.

In autonomous driving scenarios, current object detection models show strong performance when tested in clear weather. However, their performance deteriorates significantly when tested in degrading weather conditions. In addition, even when adapted to perform robustly in a sequence of different weather conditions, they are often unable to perform well in all of them and suffer from catastrophic forgetting. To efficiently mitigate forgetting, we propose Domain-Incremental Learning through Activation Matching (DILAM), which employs unsupervised feature alignment to adapt only the affine parameters of a clear weather pre-trained network to different weather conditions. We propose to store these affine parameters as a memory bank for each weather condition and plug-in their weather-specific parameters during driving (i.e. test time) when the respective weather conditions are encountered. Our memory bank is extremely lightweight, since affine parameters account for less than 2% of a typical object detector. Furthermore, contrary to previous domain-incremental learning approaches, we do not require the weather label when testing and propose to automatically infer the weather condition by a majority voting linear classifier.

Recently, large-scale pre-trained Vision and Language (VL) models have set a new state-of-the-art (SOTA) in zero-shot visual classification enabling open-vocabulary recognition of potentially unlimited set of categories defined as simple language prompts. However, despite these great advances, the performance of these zero-shot classifiers still falls short of the results of dedicated (closed category set) classifiers trained with supervised fine-tuning. In this paper we show, for the first time, how to reduce this gap without any labels and without any paired VL data, using an unlabeled image collection and a set of texts auto-generated using a Large Language Model (LLM) describing the categories of interest and effectively substituting labeled visual instances of those categories. Using our label-free approach, we are able to attain significant performance improvements over the zero-shot performance of the base VL model and other contemporary methods and baselines on a wide variety of datasets, demonstrating absolute improvement of up to $11.7\%$ ($3.8\%$ on average) in the label-free setting. Moreover, despite our approach being label-free, we observe $1.3\%$ average gains over leading few-shot prompting baselines that do use 5-shot supervision.

LiDAR 3D object detection models are inevitably biased towards their training dataset. The detector clearly exhibits this bias when employed on a target dataset, particularly towards object sizes. However, object sizes vary heavily between domains due to, for instance, different labeling policies or geographical locations. State-of-the-art unsupervised domain adaptation approaches outsource methods to overcome the object size bias. Mainstream size adaptation approaches exploit target domain statistics, contradicting the original unsupervised assumption. Our novel unsupervised anchor calibration method addresses this limitation. Given a model trained on the source data, we estimate the optimal target anchors in a completely unsupervised manner. The main idea stems from an intuitive observation: by varying the anchor sizes for the target domain, we inevitably introduce noise or even remove valuable object cues. The latent object representation, perturbed by the anchor size, is closest to the learned source features only under the optimal target anchors. We leverage this observation for anchor size optimization. Our experimental results show that, without any retraining, we achieve competitive results even compared to state-of-the-art weakly-supervised size adaptation approaches. In addition, our anchor calibration can be combined with such existing methods, making them completely unsupervised.

Pedestrians and cyclists suffer the most serious injuries in traffic accidents. Existing Pedestrian Protection Systems and Road Safety Systems rely on an ideal model of pedestrian behavior and do not consider that people tend to take shortcuts, appear at unexpected places or can be distracted on the road, for example, by using a smartphone or wearing headphones. Collecting and analyzing realistic road user behavior is a crucial component to improve pedestrian and cyclist safety. However, such real-world data is still missing. To address this, we propose a visual surveillance system with two perpendicular partially overlapping fields of view, combined with a fully automated deep learning-based pipeline to process and collect video observations, detect and extract road user trajectories in real-world coordinates and estimate human attributes, such as age, gender, smartphone usage, etc. We demonstrate our prototype by deploying it in two locations in a European city.

Domain adaptation is crucial to adapt a learned model to new scenarios, such as domain shifts or changing data distributions. Current approaches usually require a large amount of labeled or unlabeled data from the shifted domain. This can be a hurdle in fields which require continuous dynamic adaptation or suffer from scarcity of data, e.g. autonomous driving in challenging weather conditions. To address this problem of continuous adaptation to distribution shifts, we propose Dynamic Unsupervised Adaptation (DUA). By continuously adapting the statistics of the batch normalization layers we modify the feature representations of the model. We show that by sequentially adapting a model with only a fraction of unlabeled data, a strong performance gain can be achieved. With even less than 1% of unlabeled data from the target domain, DUA already achieves competitive results to strong baselines. In addition, the computational overhead is minimal in contrast to previous approaches. Our approach is simple, yet effective and can be applied to any architecture which uses batch normalization as one of its components. We show the utility of DUA by evaluating it on a variety of domain adaptation datasets and tasks including object recognition, digit recognition and object detection.

While 3D object detection in LiDAR point clouds is well-established in academia and industry, the explainability of these models is a largely unexplored field. In this paper, we propose a method to generate attribution maps for the detected objects in order to better understand the behavior of such models. These maps indicate the importance of each 3D point in predicting the specific objects. Our method works with black-box models: We do not require any prior knowledge of the architecture nor access to the model's internals, like parameters, activations or gradients. Our efficient perturbation-based approach empirically estimates the importance of each point by testing the model with randomly generated subsets of the input point cloud. Our sub-sampling strategy takes into account the special characteristics of LiDAR data, such as the depth-dependent point density. We show a detailed evaluation of the attribution maps and demonstrate that they are interpretable and highly informative. Furthermore, we compare the attribution maps of recent 3D object detection architectures to provide insights into their decision-making processes.

Most modern hand pose estimation methods rely on Convolutional Neural Networks (CNNs), which typically require a large training dataset to perform well. Exploiting unlabeled data provides a way to reduce the required amount of annotated data. We propose to take advantage of a geometry-aware representation of the human hand, which we learn from multiview images without annotations. The objective for learning this representation is simply based on learning to predict a different view. Our results show that using this objective yields clearly superior pose estimation results compared to directly mapping an input image to the 3Djoint locations of the hand if the amount of 3D annotations is limited. We further show the effect of the objective for either case, using the objective for pre-learning as well as to simultaneously learn to predict novel views and to estimate the 3D pose of the hand.

We introduce a simple yet effective fusion method of LiDAR and RGB data to segment LiDAR point clouds. Utilizing the dense native range representation of a LiDAR sensor and the setup calibration, we establish point correspondences between the two input modalities. Subsequently, we are able to warp and fuse the features from one domain into the other. Therefore, we can jointly exploit information from both data sources within one single network. To show the merit of our method, we extend SqueezeSeg, a point cloud segmentation network, with an RGB feature branch and fuse it into the original structure. Our extension called FuseSeg leads to an improvement of up to 18% IoU on the KITTI benchmark. In addition to the improved accuracy, we also achieve real-time performance at 50 fps, five times as fast as the recording speed of the KITTI LiDAR data.

Learning similarity functions between image pairs with deep neural networks yields highly correlated activations of embeddings. In this work, we show how to improve the robustness of such embeddings by exploiting the independence within ensembles. To this end, we divide the last embedding layer of a deep network into an embedding ensemble and formulate the task of training this ensemble as an online gradient boosting problem. Each learner receives a reweighted training sample from the previous learners. Further, we propose two loss functions which increase the diversity in our ensemble. These loss functions can be applied either for weight initialization or during training. Together, our contributions leverage large embedding sizes more effectively by significantly reducing correlation of the embedding and consequently increase retrieval accuracy of the embedding. Our method works with any differentiable loss function and does not introduce any additional parameters during test time. We evaluate our metric learning method on image retrieval tasks and show that it improves over state-of-the-art methods on the CUB-200-2011, Cars-196, Stanford Online Products, In-Shop Clothes Retrieval and VehicleID datasets. Therefore, our findings suggest that by dividing deep networks at the end into several smaller and diverse networks, we can significantly reduce overfitting.

When the number of categories is growing into thousands, large-scale image retrieval becomes an increasingly hard task. Retrieval accuracy can be improved by learning distance metric methods that separate categories in a transformed embedding space. Unlike most methods that utilize a single embedding to learn a distance metric, we build on the idea of boosted metric learning, where an embedding is split into a boosted ensemble of embeddings. While in general metric learning is directly applied on fine labels to learn embeddings, we take this one step further and incorporate hierarchical label information into the boosting framework and show how to properly adapt loss functions for this purpose. We show that by introducing several sub-embeddings which focus on specific hierarchical classes, the retrieval accuracy can be improved compared to standard flat label embeddings. The proposed method is especially suitable for exploiting hierarchical datasets or when additional labels can be retrieved without much effort. Our approach improves R@1 over state-of-the-art methods on the biggest available retrieval dataset (Stanford Online Products) and sets new reference baselines for hierarchical metric learning on several other datasets (CUB-200-2011, VegFru, FruitVeg-81). We show that the clustering quality in terms of NMI score is superior to previous works.

We propose a fully automated, vision-based traffic light for pedestrians. Traditional industrial solutions only report people standing in a constrained waiting zone near the crosswalk. However, reporting only people below the traffic light does not allow for efficient traffic scheduling. For example, some pedestrians do not want to cross the street and walk past the traffic light, or just wait for another person to arrive. In contrast, our system leverages intent prediction to estimate which pedestrians are actually going to cross the road by analyzing both short-term and long-term trajectory cues. In this way, we can decrease the waiting times and pave the road for optimal and adaptive traffic light scheduling. We conduct a long-term evaluation in a European capital that proves the applicability and reliability of our system and demonstrates that it is not only able to replace existing push-button solutions but also yields additional information that can be used to further optimize traffic light scheduling.

We present a novel approach for spatiotemporal saliency detection by optimizing a unified criterion of color contrast, motion contrast, appearance, and background cues. To this end, we first abstract the video by temporal superpixels. Second, we propose a novel graph structure exploiting the saliency cues to assign the edge weights. The salient segments are then extracted by applying a spectral foreground detection method, quantum cuts, on this graph. We evaluate our approach on several public datasets for video saliency and activity localization to demonstrate the favorable performance of the proposed video quantum cuts compared to the state of the art.

Learning similarity functions between image pairs with deep neural networks yields highly correlated activations of large embeddings. In this work, we show how to improve the robustness of embeddings by exploiting independence in ensembles. We divide the last embedding layer of a deep network into an embedding ensemble and formulate training this ensemble as an online gradient boosting problem. Each learner receives a reweighted training sample from the previous learners. This leverages large embedding sizes more effectively by significantly reducing correlation of the embedding and consequently increases retrieval accuracy of the embedding. Our method does not introduce any additional parameters and works with any differentiable loss function. We evaluate our metric learning method on image retrieval tasks and show that it improves over state-ofthe- art methods on the CUB-200-2011, Cars-196, Stanford Online Products, In-Shop Clothes Retrieval and VehicleID datasets by a significant margin.

We present a novel video saliency detection method to support human activity recognition and weakly supervised training of activity detection algorithms. Recent research has emphasized the need for analyzing salient information in videos to minimize dataset bias or to supervise weakly labeled training of activity detectors. In contrast to previous methods we do not rely on training information given by either eye-gaze or annotation data, but propose a fully unsupervised algorithm to find salient regions within videos. In general, we enforce the Gestalt principle of figure-ground segregation for both appearance and motion cues. We introduce an encoding approach that allows for efficient computation of saliency by approximating joint feature distributions. We evaluate our approach on several datasets, including challenging scenarios with cluttered background and camera motion, as well as salient object detection in images. Overall, we demonstrate favorable performance compared to state-of-the-art methods in estimating both ground-truth eye-gaze and activity annotations.

In this paper, we address the problem of model-free online object tracking based on color representations. According to the findings of recent benchmark evaluations, such trackers often tend to drift towards regions which exhibit a similar appearance compared to the object of interest. To overcome this limitation, we propose an efficient discriminative object model which allows us to identify potentially distracting regions in advance. Furthermore, we exploit this knowledge to adapt the object representation beforehand so that distractors are suppressed and the risk of drifting is significantly reduced. We evaluate our approach on recent online tracking benchmark datasets demonstrating state-of-the-art results. In particular, our approach performs favorably both in terms of accuracy and robustness compared to recent tracking algorithms. Moreover, the proposed approach allows for an efficient implementation to enable online object tracking in real-time.

Robust multi-object tracking-by-detection requires the correct assignment of noisy detection results to object trajectories. We address this problem by proposing an online approach based on the observation that object detectors primarily fail if objects are significantly occluded. In contrast to most existing work, we only rely on geometric information to efficiently overcome detection failures. In particular, we exploit the spatio-temporal evolution of occlusion regions, detector reliability, and target motion prediction to robustly handle missed detections. In combination with a conservative association scheme for visible objects, this allows for real-time tracking of multiple objects from a single static camera, even in complex scenarios. Our evaluations on publicly available multi-object tracking benchmark datasets demonstrate favorable performance compared to the state-of-the-art in online and offline multi-object tracking.

Combining foreground images from multiple views by projecting them onto a common ground-plane has been recently applied within many multi-object tracking approaches. These planar projections introduce severe artifacts and constrain most approaches to objects moving on a common 2D ground-plane. To overcome these limitations, we introduce the concept of an occupancy volume - exploiting the full geometry and the objects' center of mass - and develop an efficient algorithm for 3D object tracking. Individual objects are tracked using the local mass density scores within a particle filter based approach, constrained by a Voronoi partitioning between nearby trackers. Our method benefits from the geometric knowledge given by the occupancy volume to robustly extract features and train classifiers on-demand, when volumetric information becomes unreliable. We evaluate our approach on several challenging real-world scenarios including the public APIDIS dataset. Experimental evaluations demonstrate significant improvements compared to state-of-the-art methods, while achieving real-time performance.