Multi-agent trajectory prediction is a key task in computer vision for autonomous systems, particularly in dense and interactive environments. Existing methods often struggle to jointly model goal-driven behavior and complex social dynamics, which leads to unrealistic predictions. In this paper, we introduce VISTA, a recursive goal-conditioned transformer architecture that features (1) a cross-attention fusion mechanism to integrate long-term goals with past trajectories, (2) a social-token attention module enabling fine-grained interaction modeling across agents, and (3) pairwise attention maps to show social influence patterns during inference. Our model enhances the single-agent goal-conditioned approach into a cohesive multi-agent forecasting framework. In addition to the standard evaluation metrics, we also consider trajectory collision rates, which capture the realism of the joint predictions. Evaluated on the high-density MADRAS benchmark and on SDD, VISTA achieves state-of-the-art accuracy with improved interaction modeling. On MADRAS, our approach reduces the average collision rate of strong baselines from 2.14% to 0.03%, and on SDD, it achieves a 0% collision rate while outperforming SOTA models in terms of ADE/FDE and minFDE. These results highlight the model’s ability to generate socially compliant, goal-aware, and interpretable trajectory predictions, making it well-suited for deployment in safety-critical autonomous systems.

In this paper, we aim to identify how various architectural choices in two-stage one-to-many unimodal LIDAR 3D detectors influence the calibration of their IoU score (SIoU) prediction branch. We select a relevant family of freely available networks, and extensively test their output scores w.r.t. SIoUregression, and score-based ordering of predicted bounding box (BBox). We observe that the interpretation of predicted localization score (ŜIoU) as a confidence score is poorly justified, that geometric features seem to be of importance when it comes to accurately predicting SIoU, and that accurate ranking over the whole score range seems to be favored by the use of a more balanced, in terms of target metric values, distribution of BBoxes in training the score.

Deep neural networks perform well in many applications but often fail when exposed to out-of-distribution (OoD) inputs. We identify a geometric phenomenon in the embedding space: in-distribution (ID) data show higher variance than OoD data under stochastic perturbations. Using high-dimensional geometry and statistics, we explain this behavior and demonstrate its application in improving OoD detection. Unlike traditional post-hoc methods, our approach integrates uncertainty-aware tools, such as Bayesian approximations, directly into the detection process. Then, we show how considering the unit hypersphere enhances the separation of ID and OoD samples. Our mathematically sound method achieves competitive performance while remaining simple.

Uncertainty quantification is critical for deploying deep neural networks (DNNs) in real-world applications. An Auxiliary Uncertainty Estimator (AuxUE) is one of the most effective means to estimate the uncertainty of the main task prediction without modifying the main task model. To be considered robust, an AuxUE must be capable of maintaining its performance and triggering higher uncertainties while encountering Out-of-Distribution (OOD) inputs, i.e., to provide robust aleatoric and epistemic uncertainty. However, for vision regression tasks, current AuxUE designs are mainly adopted for aleatoric uncertainty estimates, and AuxUE robustness has not been explored. In this work, we propose a generalized AuxUE scheme for more robust uncertainty quantification on regression tasks. Concretely, to achieve a more robust aleatoric uncertainty estimation, different distribution assumptions are considered for heteroscedastic noise, and Laplace distribution is finally chosen to approximate the prediction error. For epistemic uncertainty, we propose a novel solution named Discretization-Induced Dirichlet pOsterior (DIDO), which models the Dirichlet posterior on the discretized prediction error. Extensive experiments on age estimation, monocular depth estimation, and super-resolution tasks show that our proposed method can provide robust uncertainty estimates in the face of noisy inputs and that it can be scalable to both image-level and pixel-wise tasks.

Trajectory prediction is a critical task for autonomous systems such as self-driving cars, surveillance systems, and social robots. The goal is to predict the future paths of road users, including cars, bikes, and pedestrians, by using their historical movement patterns and the surrounding environment. While traditional models based on Newton’s laws and social interaction forces have been used in the past, data-driven methods have become essential for learning complex spatial and temporal interactions between pedestrians. In this work, we propose SwYn-Net, a model for predicting long-term trajectories of pedestrians that can handle the uncertainty of multiple plausible paths and address the accumulation of errors over time. Our approach uses a shifted-window attention mechanism to capture the scene’s local and global context. We evaluate our model on the SDD, inD and the MOT20 datasets and show that SwYn-Net can be an efficient model to predict short-term and long-term trajectories.

The distribution of modern deep neural networks (DNNs) weights -- crucial for uncertainty quantification and robustness -- is an eminently complex object due to its extremely high dimensionality. This paper presents one of the first large-scale explorations of the posterior distribution of deep Bayesian Neural Networks (BNNs), expanding its study to real-world vision tasks and architectures. Specifically, we investigate the optimal approach for approximating the posterior, analyze the connection between posterior quality and uncertainty quantification, delve into the impact of modes on the posterior, and explore methods for visualizing the posterior. Moreover, we uncover weight-space symmetries as a critical aspect for understanding the posterior. To this extent, we develop an in-depth assessment of the impact of both permutation and scaling symmetries that tend to obfuscate the Bayesian posterior. While the first type of transformation is known for duplicating modes, we explore the relationship between the latter and L2 regularization, challenging previous misconceptions. Finally, to help the community improve our understanding of the Bayesian posterior, we release the first large-scale checkpoint dataset, including thousands of real-world models, along with our code.

Bayesian Neural Networks (BNNs) have long been considered an ideal, yet unscalable solution for improving the robustness and the predictive uncertainty of deep neural networks. While they could capture more accurately the posterior distribution of the network parameters, most BNN approaches are either limited to small networks or rely on constraining assumptions, e.g., parameter independence. These drawbacks have enabled prominence of simple, but computationally heavy approaches such as Deep Ensembles, whose training and testing costs increase linearly with the number of networks. In this work we aim for efficient deep BNNs amenable to complex computer vision architectures, e.g., ResNet-50 DeepLabv3+, and tasks, e.g., semantic segmentation and image classification, with fewer assumptions on the parameters. We achieve this by leveraging variational autoencoders (VAEs) to learn the interaction and the latent distribution of the parameters at each network layer. Our approach, called Latent-Posterior BNN (LP-BNN), is compatible with the recent BatchEnsemble method, leading to highly efficient (in terms of computation and memory during both training and testing) ensembles. LP-BNNs attain competitive results across multiple metrics in several challenging benchmarks for image classification, semantic segmentation, and out-of-distribution detection.

Monocular depth is important in many tasks, such as 3D reconstruction and autonomous driving. Deep learning based models achieve state-of-the-art performance in this field. A set of novel approaches for estimating monocular depth consists of transforming the regression task into a classification one. However, there is a lack of detailed descriptions and comparisons for Classification Approaches for Regression (CAR) in the community and no in-depth exploration of their potential for uncertainty estimation. To this end, this paper will introduce a taxonomy and summary of CAR approaches, a new uncertainty estimation solution for CAR, and a set of experiments on depth accuracy and uncertainty quantification for CAR-based models on KITTI dataset. The experiments reflect the differences in the portability of various CAR methods on two backbones. Meanwhile, the newly proposed method for uncertainty estimation can outperform the ensembling method with only one forward propagation.

This work deals with epipole estimation related to egocentric camera localization in surveillance and security applications. Matching visual features in the images provides some evidences for various solutions, so that epipole localization can be addressed as a fusion task with a large number of sources including outlier ones. In order to deal with source imprecision and uncertainty, we rely on the belief function theory and a 2D framework suited for our application. In this framework, we address the challenges introduced by a large number of sources with a strategy based on clustering and intra-cluster fusion. The proposed method exhibits more robustness in terms of accuracy and precision when compared on real data with the standard algorithms which provide single solution. Since we provide a Basic Belief Assignment as a result, our strategy is particularly adapted for the prospective combination with additional sources of information.

Tracking an entire high-density crowd composed of more than five hundred individuals is a difficult task that has not yet been accomplished. In this article, we propose to track pedestrians using a model composed of a Particle Filter (PF) and three Deep Convolutional Neural Networks (DCNN). The first network is a detector that learns to localize the persons. The second one is a pretrained network that estimates the optical flow, and the last one corrects the flow. Our contribution resides in the way we train this last network by PF supervision, and in Markov Random Field linking the different tracks.

Preventive conservation is an important practice in Cultural Heritage. The constant monitoring of the state of conservation of an artwork helps us reduce the risk of damage and number of necessary interventions. In this work, we propose a probabilistic approach for the detection of alterations on the surface of historical violins based on an a-contrario framework. Our method is a one step NFA clustering solution which considers grey-level and spatial density information in one background model. The proposed method is robust to noise and avoids parameter tuning and any assumption about the quantity of the worn-out areas. We have used as input UV induced fluorescence (UVIFL) images for considering details not perceivable with visible light. Tests were conducted on image sequences included in the “Violins UVIFL imagery” dataset. Results illustrate the ability of the algorithm to distinguish the worn area from the surrounding regions. Comparisons with state-of-the-art clustering methods show improved overall precision and recall.

A prerequisite for using smart camera networks effectively is a precise extrinsic calibration of the camera sensors, either in a fixed coordinate system, or relatively to each other. For cameras with partly overlapping fields of view, the relative pose estimation may be directly performed on or assisted by the video content obtained during scene analysis. In typical conditions however (wide baseline, repetitive patterns, homogeneous appearance of pedestrians), the pose estimation is imprecise and very often is affected by large errors in weakly constrained areas of the field of view. In this work, we propose to rely on progressively stricter constraints on the feature association between the camera views, guided by a pedestrian detector and a re-identification algorithm respectively. The results show that the two strategies are effective in alleviating the ambiguity which is due to the similar appearance of pedestrians in such scenes, and in improving the relative pose estimation.

In urban cluttered scenes, a photo provided by a wearable camera may be used by a walking law-enforcement agent as an additional source of information for localizing themselves, or elements of interest related to public safety and security. In this work, we study the problem of locating the epipole, corresponding to the position of the moving camera, in the field of view of a reference camera. We show that the presence of outliers in the standard pipeline for camera relative pose estimation not only prevents the correct estimation of the epipole localization but also degrades the standard uncertainty propagation for the epipole position. We propose a robust method for constructing an epipole location map, and we evaluate its accuracy as well as its level of integrity with respect to standard approaches.

In this work, we perform an in-depth analysis of the specific difficulties a crowded scene dataset raises for tracking algorithms. Starting from the standard characteristics depicting the crowd and their limitations, we introduce six entropy measures related to the motion patterns and to the appearance variability of the individuals forming the crowd, and one appearance measure based on Principal Component Analysis. The proposed measures are discussed on synthetic configurations and on multiple real datasets. These criteria are able to characterize the crowd behavior at a more detailed level and may be helpful for evaluating the tracking difficulty of different datasets. The results are in agreement with the perceived difficulty of the scenes.

In this work, we use the Belief Function Theory which extends the probabilistic framework in order to provide uncertainty bounds to different categories of crowd density estimators. Our method allows us to compare the multi-scale performance of the estimators, and also to characterize their reliability for crowd monitoring applications requiring varying degrees of prudence.

Detection and tracking of pedestrians in vast crowded areas is a complex problem addressed actively by the computer vision community. Proposed algorithms should ideally tackle issues of accuracy and speed at the same time. Lengthy computation times for high-quality optimization-based algorithms relying on multiple sensors make them impractical to use on long and detailed sequences. Hence, an efficient acceleration scheme, which preserves the overall accuracy, is vital to be considered. In the current work, we iterate various steps taken to accelerate a multi-camera pedestrian detection algorithm formulated as an optimization of a height map with local scene geometry constraints. The work is performed using the NVIDIA CUDA framework which allows us to efficiently utilize GPU processors and optimize the various memory accesses. The final results show more than 1000x speedup on real data frames. With respect to preserving the output accuracy, we achieve an accelerated output which is more than 99.9% in agreement with the original results.

Ahstract- This paper introduces an innovative approach for handling 2D compound hypotheses within the Belief Function Theory framework. We propose a polygon-based generic representation which relies on polygon clipping operators. This approach allows us to account in the computational cost for the precision of the representation independently of the cardinality of the discernment frame. For the BBA combination and decision making, we propose efficient algorithms which rely on hashes for fast lookup, and on a topological ordering of the focal elements within a directed acyclic graph encoding their interconnections. Additionally, an implementation of the functionalities proposed in this paper is provided as an open source library. Experimental results on a pedestrian localization problem are reported. The experiments show that the solution is accurate and that it fully benefits from the scalability of the 2D search space granularity provided by our representation.

Large scale social events are characterized by very high densities (at least locally) and an increased risk of congestions and fatal accidents. Our work focuses on the specific problem of pedestrian detection in high-density crowd images, denoted by strong homogeneity and clutter. We propose and compare different evidential fusion algorithms which are able to exploit multiple detectors based on different gradient, texture and orientation descriptors. The evidential framework allows us to model spatial imprecision arising from each of the detectors, both in the calibration and in the spatial domains. Moreover, we propose a Belief Function allocation that takes into account both types of imprecision. Results on difficult high-density crowd images acquired at Makkah during the Muslim pilgrimage show that the proposed combined fusion algorithm leads to better results than taking into account only individual sources of imprecision.

This paper addresses the problem of pedestrian detection in high-density crowd images, characterized by strong homogeneity and clutter. We propose an evidential fusion algorithm which is able to exploit multiple detectors based on different gradient, texture and orientation descriptors. The evidential framework allows us to model the spatial imprecision arising from each of the detectors. A first result of our study is that the fusion results underline clearly the good complementarity among the four descriptors we considered for this specific context. Moreover, the proposed algorithm outperforms a fusion solution based on Multiple Kernel Learning on difficult high-density crowd images acquired at Makkah at the height of the Muslim pilgrimage.

Efficient crowd counting is an essential task in crowd monitoring, and significant advances have been made in this field recently by counting-by-regression techniques. We propose in this work a learning-to-count strategy with a generic detection algorithm which benefits from a counting regressor in order to identify crowded subregions with inadequate head detection performance, and to improve their representativeness in the training set. A straightforward but crucial step is proposed in order to take into account perspective correction within the proposed framework. An evaluation on Makkah images with medium to very high densities demonstrates the effectiveness of our algorithm and its capability to reach a count error of less than 5% in this difficult setting.

Despite many proposed solutions, multi-object tracking remains a challenging problem in complex situations involving partial occlusions and non-uniform and abrupt illumination changes. Considering modular systems, the tracking performance strongly depends on the consistency of the different blocks relatively to error features. In this work, using the Belief Function framework, we take into account the reliability and the imprecision of the object detection and location to characterize objects and to derive a reliable descriptor. Since this latter is then estimated only on safe object subparts, even in case of crosses between objects, we use a distance between descriptor robust to partial occlusion, namely the recently proposed Bin-Ratio-Distance. Results obtained on various actual sequences underline the interest of the proposed algorithm by outperforming the tested alternative approaches.

Feature matching plays an important role in many computer vision applications, such as object recognition, scene reconstruction or image mosaicing. In this paper, we propose an algorithm called Hessian ORB - Overlapped FREAK (HOOFR) which is based on the combination of the ORB detector and the FREAK bio-inspired descriptor. We address some modifications related to the detection and the description processes in order to enhance HOOFR reliability, speed and memory fingerprint. The experiments on a widely used dataset demonstrate the considerable performance of HOOFR compared to SIFT, SURF or ORB in terms of the execution time and the matching quality, in various matching contexts.

Robust wide baseline pose estimation is an essential step in the deployment of smart camera networks. In this work, we highlight some current limitations of conventional strategies for relative pose estimation in difficult urban scenes. Then we propose a solution which relies on an adaptive search of corresponding interest points in synchronized video streams which allows us to converge robustly towards a high-quality solution. The experiments are performed using a manually annotated ground truth of a large scale scene exhibiting significant depth and perspective variation, uniform areas, repetitive patterns and homogeneous dynamic elements. The results show a fast and robust convergence of the solution, and a significant improvement, compared to single image based alternatives, of the RMSE of ground truth matches, and of the maximum absolute error.

This paper studies the problem of crack detection in images characterized by high gradient backgrounds. We propose an extension of a Marked Point Process model which has been successfully used for wrinkle detection. We show that our method exhibits state of the art results on a difficult image dataset, by proposing a robust trade-off between local analysis approaches, which exploit a limited amount of information around the area of interest, and global reconnection strategies, which aim to detect the crack at image level. Additional tests on a standard dataset show that the proposed method exhibits excellent performance on images with a more uniform background as well, underlining its usefulness in varying contexts.

In this study, we propose a novel solution to regulate the amount of interest points extracted from an image without significant additional computational cost. Our method acts at the very beginning of the detection process by using a corner occurrence model in order to predict the optimal threshold for a user-defined number of detections. Compared to existing approaches which guarantee a reasonable amount of corners by using a low threshold and then pruning the result, our approach is faster and more regular in terms of computation time as it avoids scoring and sorting the detected corners. Using the FAST detector as testbed, the strategy outlined in this article is evaluated in typical environments for robotics applications, and we report improved detection reliability during important scene variations. Taking into account the underlying visual navigation algorithms, we show that by regularizing the data input our solution facilitates a stable processing load, lower inter-frame computation time, and robustness to scene variations.