Julian F. P. Kooij

Shiming Wang, Holger Caesar, Liangliang Nan et al.

Multi-sensor object detection is an active research topic in automated driving, but the robustness of such detection models against missing sensor input (modality missing), e.g., due to a sudden sensor failure, is a critical problem which remains under-studied. In this work, we propose UniBEV, an end-to-end multi-modal 3D object detection framework designed for robustness against missing modalities: UniBEV can operate on LiDAR plus camera input, but also on LiDAR-only or camera-only input without retraining. To facilitate its detector head to handle different input combinations, UniBEV aims to create well-aligned Bird's Eye View (BEV) feature maps from each available modality. Unlike prior BEV-based multi-modal detection methods, all sensor modalities follow a uniform approach to resample features from the native sensor coordinate systems to the BEV features. We furthermore investigate the robustness of various fusion strategies w.r.t. missing modalities: the commonly used feature concatenation, but also channel-wise averaging, and a generalization to weighted averaging termed Channel Normalized Weights. To validate its effectiveness, we compare UniBEV to state-of-the-art BEVFusion and MetaBEV on nuScenes over all sensor input combinations. In this setting, UniBEV achieves $52.5 \%$ mAP on average over all input combinations, significantly improving over the baselines ($43.5 \%$ mAP on average for BEVFusion, $48.7 \%$ mAP on average for MetaBEV). An ablation study shows the robustness benefits of fusing by weighted averaging over regular concatenation, and of sharing queries between the BEV encoders of each modality. Our code is available at https://github.com/tudelft-iv/UniBEV.

Guopeng Li, Moritz A. Zanger, Matthijs T. J. Spaan et al.

Safe robot control requires maximizing return while satisfying safety constraints. In off-policy safe reinforcement learning, reward and safety Q-values are commonly learned by separate critic ensembles, with uncertainty handled independently for each objective. This objective-wise treatment neglects inter-objective correlation and can lead to overly conservative value estimates, thereby reducing sample efficiency. To address this issue, we propose Cholesky-Ordered Projection Q-learning (COP-Q), a safety-first method that incorporates inter-objective covariance into vector-valued Q-value estimation. COP-Q constructs a generalized confidence bound in the joint Q-value space and uses Cholesky factorization to encode objective priority in a sequential form. This preserves conservatism on safety while adaptively reducing excessive conservatism on the reward objective. The resulting estimate is used in both temporal-difference target computation and actor optimization. COP-Q incurs minimal computational overhead and is readily compatible with most existing deep Q-learning frameworks. Experiments on robot locomotion in Brax and safe navigation in Safety-Gymnasium, covering both hard- and soft-safety settings, demonstrate that COP-Q achieves strong safety performance together with competitive or improved sample efficiency relative to representative baselines.

Zimin Xia, Olaf Booij, Marco Manfredi et al.

This work addresses visual cross-view metric localization for outdoor robotics. Given a ground-level color image and a satellite patch that contains the local surroundings, the task is to identify the location of the ground camera within the satellite patch. Related work addressed this task for range-sensors (LiDAR, Radar), but for vision, only as a secondary regression step after an initial cross-view image retrieval step. Since the local satellite patch could also be retrieved through any rough localization prior (e.g. from GPS/GNSS, temporal filtering), we drop the image retrieval objective and focus on the metric localization only. We devise a novel network architecture with denser satellite descriptors, similarity matching at the bottleneck (rather than at the output as in image retrieval), and a dense spatial distribution as output to capture multi-modal localization ambiguities. We compare against a state-of-the-art regression baseline that uses global image descriptors. Quantitative and qualitative experimental results on the recently proposed VIGOR and the Oxford RobotCar datasets validate our design. The produced probabilities are correlated with localization accuracy, and can even be used to roughly estimate the ground camera's heading when its orientation is unknown. Overall, our method reduces the median metric localization error by 51%, 37%, and 28% compared to the state-of-the-art when generalizing respectively in the same area, across areas, and across time.

Ted Lentsch, Zimin Xia, Holger Caesar et al.

This work addresses cross-view camera pose estimation, i.e., determining the 3-Degrees-of-Freedom camera pose of a given ground-level image w.r.t. an aerial image of the local area. We propose SliceMatch, which consists of ground and aerial feature extractors, feature aggregators, and a pose predictor. The feature extractors extract dense features from the ground and aerial images. Given a set of candidate camera poses, the feature aggregators construct a single ground descriptor and a set of pose-dependent aerial descriptors. Notably, our novel aerial feature aggregator has a cross-view attention module for ground-view guided aerial feature selection and utilizes the geometric projection of the ground camera's viewing frustum on the aerial image to pool features. The efficient construction of aerial descriptors is achieved using precomputed masks. SliceMatch is trained using contrastive learning and pose estimation is formulated as a similarity comparison between the ground descriptor and the aerial descriptors. Compared to the state-of-the-art, SliceMatch achieves a 19% lower median localization error on the VIGOR benchmark using the same VGG16 backbone at 150 frames per second, and a 50% lower error when using a ResNet50 backbone.

Zimin Xia, Olaf Booij, Julian F. P. Kooij

We propose a novel end-to-end method for cross-view pose estimation. Given a ground-level query image and an aerial image that covers the query's local neighborhood, the 3 Degrees-of-Freedom camera pose of the query is estimated by matching its image descriptor to descriptors of local regions within the aerial image. The orientation-aware descriptors are obtained by using a translationally equivariant convolutional ground image encoder and contrastive learning. The Localization Decoder produces a dense probability distribution in a coarse-to-fine manner with a novel Localization Matching Upsampling module. A smaller Orientation Decoder produces a vector field to condition the orientation estimate on the localization. Our method is validated on the VIGOR and KITTI datasets, where it surpasses the state-of-the-art baseline by 72% and 36% in median localization error for comparable orientation estimation accuracy. The predicted probability distribution can represent localization ambiguity, and enables rejecting possible erroneous predictions. Without re-training, the model can infer on ground images with different field of views and utilize orientation priors if available. On the Oxford RobotCar dataset, our method can reliably estimate the ego-vehicle's pose over time, achieving a median localization error under 1 meter and a median orientation error of around 1 degree at 14 FPS.

Thomas M. Hehn, Julian F. P. Kooij, Dariu M. Gavrila

Various state-of-the-art self-supervised visual representation learning approaches take advantage of data from multiple sensors by aligning the feature representations across views and/or modalities. In this work, we investigate how aligning representations affects the visual features obtained from cross-view and cross-modal contrastive learning on images and point clouds. On five real-world datasets and on five tasks, we train and evaluate 108 models based on four pretraining variations. We find that cross-modal representation alignment discards complementary visual information, such as color and texture, and instead emphasizes redundant depth cues. The depth cues obtained from pretraining improve downstream depth prediction performance. Also overall, cross-modal alignment leads to more robust encoders than pre-training by cross-view alignment, especially on depth prediction, instance segmentation, and object detection.

Guopeng Li, Matthijs T. J. Spaan, Julian F. P. Kooij

When safety is formulated as a limit of cumulative cost, safe reinforcement learning (RL) aims to learn policies that maximize return subject to the cost constraint in data collection and deployment. Off-policy safe RL methods, although offering high sample efficiency, suffer from constraint violations due to cost-agnostic exploration and estimation bias in cumulative cost. To address this issue, we propose Constrained Optimistic eXploration Q-learning (COX-Q), an off-policy safe RL algorithm that integrates cost-bounded online exploration and conservative offline distributional value learning. First, we introduce a novel cost-constrained optimistic exploration strategy that resolves gradient conflicts between reward and cost in the action space and adaptively adjusts the trust region to control the training cost. Second, we adopt truncated quantile critics to stabilize the cost value learning. Quantile critics also quantify epistemic uncertainty to guide exploration. Experiments on safe velocity, safe navigation, and autonomous driving tasks demonstrate that COX-Q achieves high sample efficiency, competitive test safety performance, and controlled data collection cost. The results highlight COX-Q as a promising RL method for safety-critical applications.

Mubariz Zaffar, Liangliang Nan, Julian F. P. Kooij

In Visual Place Recognition (VPR) the pose of a query image is estimated by comparing the image to a map of reference images with known reference poses. As is typical for image retrieval problems, a feature extractor maps the query and reference images to a feature space, where a nearest neighbor search is then performed. However, till recently little attention has been given to quantifying the confidence that a retrieved reference image is a correct match. Highly certain but incorrect retrieval can lead to catastrophic failure of VPR-based localization pipelines. This work compares for the first time the main approaches for estimating the image-matching uncertainty, including the traditional retrieval-based uncertainty estimation, more recent data-driven aleatoric uncertainty estimation, and the compute-intensive geometric verification. We further formulate a simple baseline method, ``SUE'', which unlike the other methods considers the freely-available poses of the reference images in the map. Our experiments reveal that a simple L2-distance between the query and reference descriptors is already a better estimate of image-matching uncertainty than current data-driven approaches. SUE outperforms the other efficient uncertainty estimation methods, and its uncertainty estimates complement the computationally expensive geometric verification approach. Future works for uncertainty estimation in VPR should consider the baselines discussed in this work.

Nail Ibrahimli, Julian F. P. Kooij, Liangliang Nan

We introduce MuVieCAST, a modular multi-view consistent style transfer network architecture that enables consistent style transfer between multiple viewpoints of the same scene. This network architecture supports both sparse and dense views, making it versatile enough to handle a wide range of multi-view image datasets. The approach consists of three modules that perform specific tasks related to style transfer, namely content preservation, image transformation, and multi-view consistency enforcement. We extensively evaluate our approach across multiple application domains including depth-map-based point cloud fusion, mesh reconstruction, and novel-view synthesis. Our experiments reveal that the proposed framework achieves an exceptional generation of stylized images, exhibiting consistent outcomes across perspectives. A user study focusing on novel-view synthesis further confirms these results, with approximately 68\% of cases participants expressing a preference for our generated outputs compared to the recent state-of-the-art method. Our modular framework is extensible and can easily be integrated with various backbone architectures, making it a flexible solution for multi-view style transfer. More results are demonstrated on our project page: muviecast.github.io

Shireen Kudukkil Manchingal, Andrew Bradley, Julian F. P. Kooij et al.

Despite AI's impressive achievements, including recent advances in generative and large language models, there remains a significant gap in the ability of AI systems to handle uncertainty and generalize beyond their training data. AI models consistently fail to make robust enough predictions when facing unfamiliar or adversarial data. Traditional machine learning approaches struggle to address this issue, due to an overemphasis on data fitting, while current uncertainty quantification approaches suffer from serious limitations. This position paper posits a paradigm shift towards epistemic artificial intelligence, emphasizing the need for models to learn from what they know while at the same time acknowledging their ignorance, using the mathematics of second-order uncertainty measures. This approach, which leverages the expressive power of such measures to efficiently manage uncertainty, offers an effective way to improve the resilience and robustness of AI systems, allowing them to better handle unpredictable real-world environments.

Shiming Wang, Holger Caesar, Liangliang Nan et al.

In autonomous driving, multi-modal perception tasks like 3D object detection typically rely on well-synchronized sensors, both at training and inference. However, despite the use of hardware- or software-based synchronization algorithms, perfect synchrony is rarely guaranteed: Sensors may operate at different frequencies, and real-world factors such as network latency, hardware failures, or processing bottlenecks often introduce time offsets between sensors. Such asynchrony degrades perception performance, especially for dynamic objects. To address this challenge, we propose AsyncBEV, a trainable lightweight and generic module to improve the robustness of 3D Birds' Eye View (BEV) object detection models against sensor asynchrony. Inspired by scene flow estimation, AsyncBEV first estimates the 2D flow from the BEV features of two different sensor modalities, taking into account the known time offset between these sensor measurements. The predicted feature flow is then used to warp and spatially align the feature maps, which we show can easily be integrated into different current BEV detector architectures (e.g., BEV grid-based and token-based). Extensive experiments demonstrate AsyncBEV improves robustness against both small and large asynchrony between LiDAR or camera sensors in both the token-based CMT and grid-based UniBEV, especially for dynamic objects. We significantly outperform the ego motion compensated CMT and UniBEV baselines, notably by $16.6$ % and $11.9$ % NDS on dynamic objects in the worst-case scenario of a $0.5 s$ time offset. Code will be released upon acceptance.

Shireen Kudukkil Manchingal, Armand Amaritei, Mihir Gohad et al.

Autonomous Vehicle (AV) perception systems have advanced rapidly in recent years, providing vehicles with the ability to accurately interpret their environment. Perception systems remain susceptible to errors caused by overly-confident predictions in the case of rare events or out-of-sample data. This study equips an autonomous vehicle with the ability to 'know when it is uncertain', using an uncertainty-aware image classifier as part of the AV software stack. Specifically, the study exploits the ability of Random-Set Neural Networks (RS-NNs) to explicitly quantify prediction uncertainty. Unlike traditional CNNs or Bayesian methods, RS-NNs predict belief functions over sets of classes, allowing the system to identify and signal uncertainty clearly in novel or ambiguous scenarios. The system is tested in a real-world autonomous racing vehicle software stack, with the RS-NN classifying the layout of the road ahead and providing the associated uncertainty of the prediction. Performance of the RS-NN under a range of road conditions is compared against traditional CNN and Bayesian neural networks, with the RS-NN achieving significantly higher accuracy and superior uncertainty calibration. This integration of RS-NNs into Robot Operating System (ROS)-based vehicle control pipeline demonstrates that predictive uncertainty can dynamically modulate vehicle speed, maintaining high-speed performance under confident predictions while proactively improving safety through speed reductions in uncertain scenarios. These results demonstrate the potential of uncertainty-aware neural networks - in particular RS-NNs - as a practical solution for safer and more robust autonomous driving.

Mubariz Zaffar, Liangliang Nan, Sebastian Scherer et al.

Given a query image, Visual Place Recognition (VPR) is the task of retrieving an image of the same place from a reference database with robustness to viewpoint and appearance changes. Recent works show that some VPR benchmarks are solved by methods using Vision-Foundation-Model backbones and trained on large-scale and diverse VPR-specific datasets. Several benchmarks remain challenging, particularly when the test environments differ significantly from the usual VPR training datasets. We propose a complementary, unexplored source of information to bridge the train-test domain gap, which can further improve the performance of State-of-the-Art (SOTA) VPR methods on such challenging benchmarks. Concretely, we identify that the test-time reference set, the "map", contains images and poses of the target domain, and must be available before the test-time query is received in several VPR applications. Therefore, we propose to perform simple Reference-Set-Finetuning (RSF) of VPR models on the map, boosting the SOTA (~2.3% increase on average for Recall@1) on these challenging datasets. Finetuned models retain generalization, and RSF works across diverse test datasets.

Zimin Xia, Yujiao Shi, Hongdong Li et al.

Given a ground-level query image and a geo-referenced aerial image that covers the query's local surroundings, fine-grained cross-view localization aims to estimate the location of the ground camera inside the aerial image. Recent works have focused on developing advanced networks trained with accurate ground truth (GT) locations of ground images. However, the trained models always suffer a performance drop when applied to images in a new target area that differs from training. In most deployment scenarios, acquiring fine GT, i.e. accurate GT locations, for target-area images to re-train the network can be expensive and sometimes infeasible. In contrast, collecting images with noisy GT with errors of tens of meters is often easy. Motivated by this, our paper focuses on improving the performance of a trained model in a new target area by leveraging only the target-area images without fine GT. We propose a weakly supervised learning approach based on knowledge self-distillation. This approach uses predictions from a pre-trained model as pseudo GT to supervise a copy of itself. Our approach includes a mode-based pseudo GT generation for reducing uncertainty in pseudo GT and an outlier filtering method to remove unreliable pseudo GT. Our approach is validated using two recent state-of-the-art models on two benchmarks. The results demonstrate that it consistently and considerably boosts the localization accuracy in the target area.

Jetze T. Schuurmans, Kim Batselier, Julian F. P. Kooij

Tensor decompositions have been successfully applied to compress neural networks. The compression algorithms using tensor decompositions commonly minimize the approximation error on the weights. Recent work assumes the approximation error on the weights is a proxy for the performance of the model to compress multiple layers and fine-tune the compressed model. Surprisingly, little research has systematically evaluated which approximation errors can be used to make choices regarding the layer, tensor decomposition method, and level of compression. To close this gap, we perform an experimental study to test if this assumption holds across different layers and types of decompositions, and what the effect of fine-tuning is. We include the approximation error on the features resulting from a compressed layer in our analysis to test if this provides a better proxy, as it explicitly takes the data into account. We find the approximation error on the weights has a positive correlation with the performance error, before as well as after fine-tuning. Basing the approximation error on the features does not improve the correlation significantly. While scaling the approximation error commonly is used to account for the different sizes of layers, the average correlation across layers is smaller than across all choices (i.e. layers, decompositions, and level of compression) before fine-tuning. When calculating the correlation across the different decompositions, the average rank correlation is larger than across all choices. This means multiple decompositions can be considered for compression and the approximation error can be used to choose between them.

Yannick Schulz, Avinash Kini Mattar, Thomas M. Hehn et al.

This work proposes to use passive acoustic perception as an additional sensing modality for intelligent vehicles. We demonstrate that approaching vehicles behind blind corners can be detected by sound before such vehicles enter in line-of-sight. We have equipped a research vehicle with a roof-mounted microphone array, and show on data collected with this sensor setup that wall reflections provide information on the presence and direction of occluded approaching vehicles. A novel method is presented to classify if and from what direction a vehicle is approaching before it is visible, using as input Direction-of-Arrival features that can be efficiently computed from the streaming microphone array data. Since the local geometry around the ego-vehicle affects the perceived patterns, we systematically study several environment types, and investigate generalization across these environments. With a static ego-vehicle, an accuracy of 0.92 is achieved on the hidden vehicle classification task. Compared to a state-of-the-art visual detector, Faster R-CNN, our pipeline achieves the same accuracy more than one second ahead, providing crucial reaction time for the situations we study. While the ego-vehicle is driving, we demonstrate positive results on acoustic detection, still achieving an accuracy of 0.84 within one environment type. We further study failure cases across environments to identify future research directions.

Andras Palffy, Jiaao Dong, Julian F. P. Kooij et al.

This letter presents a novel radar based, single-frame, multi-class detection method for moving road users (pedestrian, cyclist, car), which utilizes low-level radar cube data. The method provides class information both on the radar target- and object-level. Radar targets are classified individually after extending the target features with a cropped block of the 3D radar cube around their positions, thereby capturing the motion of moving parts in the local velocity distribution. A Convolutional Neural Network (CNN) is proposed for this classification step. Afterwards, object proposals are generated with a clustering step, which not only considers the radar targets' positions and velocities, but their calculated class scores as well. In experiments on a real-life dataset we demonstrate that our method outperforms the state-of-the-art methods both target- and object-wise by reaching an average of 0.70 (baseline: 0.68) target-wise and 0.56 (baseline: 0.48) object-wise F1 score. Furthermore, we examine the importance of the used features in an ablation study.

Frank Hafner, Amran Bhuiyan, Julian F. P. Kooij et al.

Person re-identification is a key challenge for surveillance across multiple sensors. Prompted by the advent of powerful deep learning models for visual recognition, and inexpensive RGB-D cameras and sensor-rich mobile robotic platforms, e.g. self-driving vehicles, we investigate the relatively unexplored problem of cross-modal re-identification of persons between RGB (color) and depth images. The considerable divergence in data distributions across different sensor modalities introduces additional challenges to the typical difficulties like distinct viewpoints, occlusions, and pose and illumination variation. While some work has investigated re-identification across RGB and infrared, we take inspiration from successes in transfer learning from RGB to depth in object detection tasks. Our main contribution is a novel method for cross-modal distillation for robust person re-identification, which learns a shared feature representation space of person's appearance in both RGB and depth images. In addition, we propose a cross-modal attention mechanism where the gating signal from one modality can dynamically activate the most discriminant CNN filters of the other modality. The proposed distillation method is compared to conventional and deep learning approaches proposed for other cross-domain re-identification tasks. Results obtained on the public BIWI and RobotPKU datasets indicate that the proposed method can significantly outperform the state-of-the-art approaches by up to 16.1% in mean Average Precision (mAP), demonstrating the benefit of the distillation paradigm. The experimental results also indicate that using cross-modal attention allows to improve recognition accuracy considerably with respect to the proposed distillation method and relevant state-of-the-art approaches.