Julian Truetsch, Felix Hauser, Christoph Stiller +1cs.CV cs.CL cs.LG cs.NE cs.RO
Understanding the composition of large-scale autonomous driving datasets is essential for safety, robustness, and reliable operation across domains. For example, domain shift between locations could lead to the operating environment being misaligned with the training data, resulting in potentially dangerous performance degradation. Yet, existing data analysis pipelines largely rely on metadata, predefined labels, or manual inspection, which provide limited semantic insight or do not scale. This paper studies set difference captioning: given two subsets of images, the goal is to produce a natural-language hypothesis describing differences between the target and reference set. Building on a two-stage formulation, we adapt the method to autonomous driving by focusing on object-centric patches derived from object detection, which simplifies aggregation and enables attribution of differences to specific object instances or categories. To evaluate this setting in-domain, we introduce a new benchmark, AD-Diff Bench. Low-concentration experiments assess the suitability of set-difference-captioning approaches to sparse, real-world differences. We restrict our experiments to open-weight models to support reproducibility and ease of deployment. The proposed benchmark and analysis provide a step towards practical, human-interpretable dataset introspection for autonomous driving datasets. Our implementation and benchmark dataset are available at https://github.com/KIT-MRT/AD-Diff
Four-dimensional (4D) Radar is a powerful sensing modality capable of detecting surrounding three-dimensional (3D) objects under diverse weather conditions and providing Doppler-based motion information. However, raw 4D Radar signals contain significant clutter from road surfaces, guardrails, and surrounding vehicles, along with multipath-induced ghost reflections and the receiver's inherent noise floor. Consequently, preprocessing algorithms designed to remove such invalid measurements often make the Radar data excessively sparse. Moreover, the Doppler measurements provided by 4D Radar describe only the radial component of an object's velocity, limiting their ability to recover the full motion state. In this paper, we introduce a stereo 4D Radar-based 3D object detection framework that exploits the geometric disparity between left and right Radars to estimate the absolute velocity of objects and achieve more robust perception through the fusion of their complementary features. The effectiveness of the proposed framework is validated on our in-house stereo 4D Radar dataset, demonstrating performance gains of 8.82 points in AP 3D and 9.0 points in AP BEV over state-of-the-art mono 4D Radar baselines. These results demonstrate that absolute velocity estimation combined with stereo geometry-aware feature fusion leads to substantial improvements in 3D object detection.
Detectors trained on closed-set annotations can miss rare moving objects outside the training taxonomy. Automotive radar provides category-independent Doppler motion cues and is less affected by adverse illumination and weather, but sparse, noisy returns hinder class-aware 3D box detection. Surface location and velocity remain useful for motion reasoning and collision avoidance when full box geometry is difficult to recover. We present the Physics-Aware Radar Transformer (PART), a fully sparse radar-only detector that predicts existence confidence, a representative surface point, and 2D ground-plane velocity for each moving-object hypothesis. Doppler-Aware Query Initialization (DAQI) replaces scene-independent learned queries with input-dependent proposals by clustering radar returns in position and velocity, easing query-object assignment in sparse scenes. Physics-Guided Cross-Attention (PGCA) incorporates radial-Doppler consistency and radar cross section (RCS) into query-point association. Uncertainty-aware supervision randomly masks ground-truth objects and assigns soft existence targets to ambiguous radar-supported queries, reducing reliance on exhaustive annotations. With only 1.1 million parameters, PART achieves a class-agnostic average precision (CA-AP) of 0.8827, a mean average surface translation error (mASTE) of 0.3188 m, and a mean average velocity error (mAVE) of 0.8084 m/s on nuScenes. It attains 0.9203 recall on rare and safety-relevant categories excluded from the standard evaluation and remains effective at night, in rain, and under severe occlusion. Inspection of apparent false positives shows that some predictions correspond to moving objects absent from the nuScenes annotations. Code and pretrained model weights will be publicly available at https://github.com/sunyinghao-uestc/PART.
Christian Löwens, Thorben Funke, Alexandru Paul Condurachecs.CV cs.LG cs.RO
End-to-end autonomous driving models plan future trajectories from raw sensor input. While earlier driving benchmarks often measured deviation from the human trajectory, current benchmarks such as NAVSIM and Bench2Drive evaluate models with richer simulation-based metrics intended to capture safe and compliant driving. A high benchmark score should reflect that a model can understand the scene in front of it and act accordingly. But how much of that score specifically comes from reacting to the dynamic part of that scene? To probe this, we remove a model's camera input and replace it with memories from prior drives at the same location. The retrieved memories can provide persistent scene information, including road layout and location-conditioned regularities, but not the current traffic state. Surprisingly, memory is nearly sufficient on NAVSIM, reaching or even exceeding the performance of leading end-to-end methods without actually observing the evaluated scene. Our results suggest that a high NAVSIM score does not require a planner to react to the current traffic scene and should be treated with caution. This effect is benchmark-dependent: driving from memory causes substantially larger performance drops on Bench2Drive and RealEngine. We provide our code at https://github.com/boschresearch/MemoryDrivoR .
High-dynamic-range (HDR) images, with their rich tone and detail reproduction, hold significant potential to enhance computer vision systems, particularly in autonomous driving. However, most neural networks for embedded systems are trained on low-dynamic-range (LDR) inputs and suffer substantial performance degradation when handling high-bit-depth HDR images due to the challenges posed by extreme dynamic ranges. In this paper, we propose a novel tone mapping method that not only bridges the gap between HDR RAW inputs and the LDR sRGB requirements of detection networks but also achieves end-to-end optimization with downstream tasks. Instead of relying on the traditional image signal processing (ISP) pipeline, we introduce neural photometric calibration to regularize dynamic ranges and a scaling-invariant local tone mapping model to preserve image details. In addition, our architecture also supports performance transfer finetuning, enabling efficient adaptation from the LDR sRGB images to the HDR RAW images with minimal cost. The proposed method outperforms traditional tone mapping algorithms and advanced AI-ISP methods in challenging automotive HDR scenes. Moreover, our pipeline achieves real-time processing of 4K high-bit-depth HDR inputs on NVIDIA Jetson platforms.
Rahul Ahuja, Bala Murali Manoghar Sai Sudhakar, Shashwata Gupta +3cs.CV
Three-dimensional box-and-track annotation is the cost bottleneck in autonomous-driving data engines, and the offline systems built to relieve it replace the online perception stack outright, so a team needing both regimes maintains and reconciles two. MotionSync makes the causal/non-causal boundary an explicit architectural seam instead. A strictly causal tracker, built on a strong published baseline and extended with innovation-driven uncertainty calibration, frame-rate-invariant kinematic association gates, and multi-hypothesis motion with learned mode selection, emits a valid online result. A non-causal pass then revises the buffered trajectories with Rauch--Tung--Striebel smoothing applied separately to pose, extent and yaw, physics-validated gap completion, and semantic pruning of ghost tracks against LiDAR point labels. The refiner never writes back, so one system serves both regimes and refinement's effect is a delta over an unaltered causal estimate. Used as an auto-labeller, a fixed 3D detector trained on 25% human labels plus MotionSync pseudo-labels reaches 96.9% of its full-supervision mean average precision (mAP) on Waymo, and at a 10% budget the non-causal pass accounts for +3.3 mAP/L2 over pseudo-labels from the same tracker's causal stage. Re-fitting the online tracker on its own refined output recovers 73% of the benefit of human supervision, while its causal output is worse supervision than no re-fitting at all. As a tracker MotionSync is at parity with the leading published offline entries on the headline metric and ahead of them on error composition, which is where a refinement pass can act at all: it reduces misses and fragmentations together, the signature of gap completion rather than of a tuned detector.
The point-based occupancy prediction paradigm has achieved an attractive trade-off between accuracy and efficiency by modeling 3D space sparsely. However, its predictions inherently mismatch the dense voxel-based occupancy required by self-driving systems, necessitating hand-crafted heuristics during training and inference that limit final performance. To overcome these limitations, we propose OPUS-V2, a novel framework built upon the pioneering OPUS (occupancy prediction using a sparse set) point-based approach. OPUS-V2 incorporates a lightweight point-voxel transformation (PVT) module behind the decoder to adaptively map sparse predictions into the dense voxel space, eliminating the need for suboptimal operations and improving model accuracy. Furthermore, our architecture decouples feature and occupancy generation processes, allowing OPUS-V2 to adapt to arbitrary occupancy resolutions. OPUS-V2 achieves a state-of-the-art rayIoU of 44.0 on the Occ3D dataset. On the more challenging OpenOccupancy dataset, it attains a competitive 16.4 mIoU while running in real time at 20.6 FPS.
Video generation for autonomous driving cannot follow the web-scale route: driving data is expensive to collect, bound by privacy requirements, and cannot be scraped at will, so models must make the most of a fixed corpus. We present a systematic scaling-law study of video diffusion models trained from scratch on driving data: a family of models from 1M to 9B parameters, trained at different exposures on up to 5,500 hours of driving. Validation loss follows consistent power laws in both model size and training exposure, answering the questions that shape a training budget: whether compute is better spent on longer training or on a larger model, and whether more data is needed. Loss improves much faster with training exposure than with model size, making longer training the most effective way to improve a fixed model under limited compute. However, larger models continue to achieve lower asymptotic loss, so compute-optimal scaling still favors increasing model size when sufficient compute and data are available. Guided by these laws, we train a 9B-parameter model, to our knowledge the largest video diffusion model trained from scratch on driving data: it sets a new open-source state of the art for driving video generation, as measured on nuScenes. Our code and pretrained models are available at https://github.com/valeoai/VATIX. NATIX is separately releasing the underlying driving data in stages.
Image-to-Point Cloud Registration aims to estimate the camera pose of a given image within a 3D scene point cloud, which is a fundamental task in autonomous driving and large-scale outdoor localization. Recent implicit correspondence learning methods have improved registration performance by learning cross-modal alignment in an end-to-end framework, leading to more accurate camera pose estimation. However, due to the inherent modality discrepancy between images and sparse LiDAR point clouds, reliable cross-modal correspondence learning remains challenging. To address this issue, we propose Depth-Guided Projective Alignment for Image-to-Point-Cloud Registration (DPA-I2P). Unlike naive depth or feature concatenation, Ray-Conditioned Metric Depth Encoding (RMDE) and Projection-Consistent Vision Lifting (PVL) exploit depth and visual cues in a structured, geometry-aware manner. In addition, Cross-Modal Query Pruning (CQP) suppresses unreliable queries during early refinement to improve matching stability. Experiments on KITTI and nuScenes demonstrate the effectiveness of the proposed method. On KITTI, DPA-I2P reduces RTE and RRE by 45.0% and 55.6% over the strongest implicit baseline, respectively. On nuScenes, DPA-I2P also improves registration accuracy over the evaluated baselines, suggesting better transferability to different driving scenes.
Current world action models (WAMs) typically operate on 2D visual data. These models can achieve exceptional visual quality, but they lack explicit spatial structure for individual objects and repeatedly process redundant background content. Although point clouds can represent the world in 3D space, they can be difficult to align and accumulate across viewpoints. In this paper, we leverage an explicit 4D Gaussian Splatting (4DGS) representation that separately models dynamic objects and the static background of a scene. For dynamic objects, we use a policy model to predict future actor actions and a world model to predict transformations of their observed Gaussian splats. The static background need not be regenerated for future states, as much of it has already been observed in past frames. This forms an object-centric world action model, which we name 4DGS-WAM. It lifts 2D observations into a persistent 4D representation so that previously observed static content can be reused during future prediction. Future-state extrapolation can then focus on modeling the evolution of dynamic objects. Experiments on KITTI-MOT evaluate short-horizon prediction and past reconstruction.
Autonomous vehicles operate in dynamic, ever-changing environments where new scenarios and edge cases constantly emerge. As a result, static learning models are inadequate for ensuring safe and reliable operation. Continuous learning is essential for adapting to these evolving conditions and maintaining robust performance across diverse real-world settings. However, autonomous vehicles generate massive streams of visual data during operation, and existing continuous learning approaches typically rely on heuristic sampling methods that fail to capture temporal dynamics, often overlooking critical learning opportunities or selecting redundant frames. In this paper, we introduce FrameScope, a temporal data valuation framework for continuous learning in autonomous vehicles. FrameScope extends neural tangent kernel theory to temporal domains, enabling principled valuation of streaming visual data. Unlike cloud-centric methods that transmit all video data for processing, our approach performs principled, local frame selection on the vehicle and queries a cloud-based oracle model only for labels of those high-value frames. Extensive experiments across multiple domain shifts show that FrameScope consistently outperforms existing methods, achieving higher sample efficiency and significantly reducing catastrophic forgetting in autonomous vehicle perception. By valuing data on the vehicle and querying only labels for selected frames, FrameScope reduces bandwidth requirements, enabling scalable operation with a lightweight cloud labeling service.
Matthias Neuwirth-Trapp, Begüm Altunbas, Jiayi Wang +4cs.CV cs.RO
Achieving 360° coverage is critical for the visual perception systems of autonomous vehicles. Fisheye cameras offer a cost-effective solution by enabling full surround coverage with as few as two sensors. However, existing multi-view fisheye datasets are limited, and synthesizing rare corner cases typically requires computationally expensive 3D simulations, hindering the training. While generative models have achieved significant success in standard perspective imagery, their application to wide-angle distortion remains unexplored. In this work, we formally introduce the novel problem of multi-view fisheye image generation conditioned on volumetric semantic representations and present two distinct methods. We first propose SyntheOcc-FE, which adapts the SyntheOcc architecture to fisheye data. While effective, this method is constrained by the scarcity of fisheye datasets, which limits its generalization. To overcome these limitations, we propose our second method, MIVIFI (multi-view fisheye), which leverages cross-domain learning with Equirectangular Projections. By bridging the gap between dataset domains using KITTI-360 fisheye images alongside nuScenes multi-view standard images, our approach enables high-fidelity manipulation of scene content. This framework enables the structural modification of semantic occupancy inputs to introduce or eliminate specific actors and facilitates the rendering of diverse meteorological conditions and illumination scenarios absent in the limited fisheye datasets. Quantitative and qualitative experiments demonstrate that our methods achieve robust photorealistic multi-view fisheye image generation and highlight the specific advantages of our cross-domain strategy for handling data scarcity.
This workshop paper comprehensively reviews the First Challenge on Unified Removal of Raindrops and Reflections. The challenge aims to address a frequently encountered practical problem in the field of autonomous driving, i.e., raindrop-reflection composite degradation on rainy days. This competition attracted 149 registered participants and received 12 valid final submissions with corresponding fact sheets, significantly contributing to the progress of unified removal of raindrops and reflections. All the methods are developed and evaluated on our real-shot RainDrop and ReFlection (RDRF) dataset. A detailed analysis of the submitted methods and corresponding results is provided in this report, which highlights effective approaches and provides interesting insights for future research.
Reliable traffic sign detection is a prerequisite for the global deployment of autonomous driving systems, where regulatory compliance and road safety depend on perceiving signs correctly across regions, ranges, and weather conditions. Despite recent progress, vision-based methods continue to face three fundamental limitations: poor cross-regional generalization due to high diversity across countries, degraded performance on small-object detection at long ranges (traffic signs occupy as little as $10{\times}10$ pixels at 200m), and fragile temporal tracking under the strongly non-linear perspective distortion that occurs as a vehicle approaches a sign. In this paper, we address the problem of robust, long-range, region-agnostic traffic sign perception by combining camera and Light Detection and Ranging (LiDAR) sensing. We present a multi-modal detection framework whose Intensity-Aware Deformable Fusion module aligns retro-reflective LiDAR cues with camera features, anchoring detection on geometric invariants rather than region-specific visual appearance. We further introduce a dual motion-model tracker that explicitly accounts for non-linear perspective transformations during vehicle approach, substantially improving temporal consistency over linear motion assumptions. Additionally, we develop a semantic attribute classification pipeline that estimates occlusion level, readability, sign embeddedness, and road relevance, providing actionable context to downstream planning. Extensive evaluation on our dataset, spanning 60+ countries and 2,500+ hours of driving data, shows that the proposed pipeline achieves an Object Miss Ratio (OMR) of 0.49% across 221,068 evaluation sequences, demonstrating globally generalizable traffic sign perception in commercial-grade autonomous driving systems.
While modern autonomous driving systems excel at perception tasks such as object detection and trajectory prediction, they lack the high-level causal reasoning required to interpret traffic accidents. In particular, determining responsibility, such as identifying who is at fault and which traffic rule was violated, remains largely unexplored in current benchmarks. To this end, we introduce CAViAR (Causal Accident Video and Incident Analysis Repository), a human-annotated dashcam benchmark comprising 2,249 real-world accident videos collected from CarCrashDataset (CCD) and Nexar. Each video is annotated with structured labels spanning environmental conditions, accident type, causal explanation, apparent At-Fault Agent, affected agent, and apparent rule-violation category. We benchmark state-of-the-art vision-language models (VLMs), including Cosmos-Reason2, Qwen3-VL, and InternVL3. Once class imbalance is accounted for with majority/random baselines and balanced metrics, perceptual competence is uneven--lighting is nearly solved, whereas weather and road-condition accuracy fall at or below the majority-class baseline---and all models degrade sharply on accident type and responsibility reasoning. Overall, CAViAR exposes a practical Perception--Reasoning Gap: current VLMs may recognize salient context, but do not reliably map visible agent actions to annotated rule-relevant responsibility categories in safety-critical driving scenarios. Code, annotation schema, prompts, and evaluation scripts are available at: https://github.com/nec-labs-ma/CAViAR
Spatial representation learning for autonomous driving aims to map raw visual signals into structured 3D scene representations, where object-centric bounding boxes and rendering-oriented 3D primitives (\eg, 3D Gaussians) serve as two distinct yet highly complementary levels for scene understanding. Existing methods typically treat dynamic reconstruction and instance-level perception as separate tasks, despite their shared goal of estimating the underlying 3D world state. As a result, dynamic reconstruction is under-constrained while 3D detection lacks geometric grounding. To address this gap, we propose USR-Drive, a unified conditional generative framework that, given only posed multi-view driving videos, jointly recovers dense dynamic geometry and instance-level object layouts within a shared scene representation. Specifically, USR-Drive represents dense Gaussian primitives and sparse 3D bounding boxes as two aligned latent token streams and jointly denoises them with a unified multi-modal diffusion Transformer. Unlike prior paradigms that use boxes as external conditions or predict them with detached modules, USR-Drive treats them as mutually constrained state variables with a Unified Positional Encoding (UPE) that aligns heterogeneous tokens within a shared metric spatiotemporal coordinate. Via such unified representation and generative framework, the two modalities reinforce each other: geometry supplies dense metric evidence for box prediction, while boxes provide instance-level structural priors that help preserve spatial consistency and reduce ambiguity in sequential 3D geometric representation. Our approach successfully delivers state-of-the-art results for both dynamic reconstruction and 3D detection on the nuScenes and VKitti datasets.
Jonel Roman, Ryan Sirjue, Peter Nguyen +3cs.CV cs.AI
Autonomous vehicles depend on fast and reliable perception systems to detect surrounding vehicles, pedestrians, cyclists, traffic signs, and other road objects in real time. This paper presents a comprehensive survey and analysis of one-stage object detectors for autonomous driving rather than an implementation of a new detection system. The survey reviews the evolution of major one-stage detectors, including YOLOv1, SSD, RetinaNet, EfficientDet, anchor-free detectors such as FCOS and CenterNet, and recent real-time models such as YOLOv10. It compares these architectures through their design choices, feature-fusion strategies, loss functions, deployment trade-offs, and reported benchmark performance. The paper also summarizes commonly used autonomous-driving datasets, evaluation metrics, open challenges, and future research directions. Overall, this survey highlights how one-stage detectors balance speed, accuracy, efficiency, and robustness, while also emphasizing the remaining gap between benchmark results and dependable real-world autonomous-driving performance.
Driving scene reconstruction and rendering, especially with 3D Gaussian Splatting, has become an important component of autonomous driving simulation. However, rendered views often degrade under extrapolated ego trajectories and scene edits, producing blurry structures, temporal flicker, and foreground-background misalignment. Existing refinement methods are commonly designed for a specific setting, such as image-level novel-view repair or object-editing correction. In this paper, we introduce SPVC, a structured and panoptic video fixing framework for cross-dataset driving scene rendering. The name summarizes four design principles. (1) Structured fixing denotes the use of explicit spatial conditions, including camera pose, 3D bounding boxes, and HD maps, to guide the repair process and reduce uncontrolled hallucination. (2) Panoptic fixing refers to correcting both background rendering artifacts, such as distorted roads, buildings, and lanes, and foreground vehicle artifacts introduced by scene editing, such as inconsistent object appearance. (3) Video fixing means that the model operates on driving sequences rather than isolated frames, allowing temporal cues to be used during artifact correction. (4) Cross-dataset fixing means that a single shared network is trained and applied across multiple driving datasets, reducing the need for dataset-specific or scene-specific fixers. Concretely, we construct paired degraded-clean training data by simulating under-constrained 3DGS rendering and foreground vehicle insertion artifacts, and train a two-stage controllable video diffusion model that first addresses video-level appearance and then refines scene layout with structured controls.
Christopher Lang, Alexander Braun, Abhinav Valadacs.CV
Video self-supervised learning through masked spatiotemporal prediction has emerged as a promising paradigm for learning feature representations from unlabeled data. However, existing methods typically rely on random masking, which indiscriminately removes regions irrespective of their semantic or temporal relevance. In ego-centric driving videos, this can weaken the pretext signal since safety-critical cues such as pedestrians, vehicles, lane boundaries, and dynamic interactions often occupy only a small portion of the frame, yet are central to downstream perception. We introduce V-JEPA4A, a domain-specialized variant of V-JEPA for autonomous driving that is pre-trained on publicly available driving videos with a novel saliency-driven masking policy. It accounts for semantically and temporally relevant context. The proposed policy preserves and predicts context according to semantic importance and temporal relevance, yielding more informative representation learning while retaining the efficiency of masked prediction. We evaluate the resulting encoders on four driving benchmarks spanning tracking, semantic segmentation, and depth estimation. The results demonstrate that V-JEPA4A reduces identity switches on BDD100k MOT by 25% over V-JEPA with random masking, achieves 73.2 mIoU on Cityscapes, and 3.75 RMSE on KITTI-2015 depth, while incurring only ~14% additional pre-training iteration overhead.
Azhar Hussian, Martin Vossiek, Vasileios Belagianniscs.CV
LiDAR scene completion is a key component of 3D perception in autonomous driving, where the scene must be completed in real time to be usable in downstream tasks. Existing approaches typically follow an initialize-and-refine paradigm, in which a coarse initialization of the scene is first constructed, then refined into complete 3D geometry. Generative models are slower because they iteratively refine random Gaussian noise into the scene, while non-generative methods perturb the partial scene with a fixed noise scale, which limits coverage of large gaps and occluded regions and requires manual recalibration for each new sensor configuration. We present RapidLiDAR, a LiDAR scene completion method that treats the initialization itself as a learned, data-driven component. We propose an adaptive initialization module that predicts a spatially varying displacement for each partial input point, expanding the partial observations into a coarse scene initialization adapted to the local geometry, without requiring manual noise tuning. To refine this coarse initialization into a complete and coherent scene, we additionally propose a multi-scale reconstruction module that further refines point positions by querying multi-scale 3D voxel and 2D BEV feature maps constructed from the input scan. By replacing point-neighborhood operators such as farthest point sampling and $k$-nearest neighbor search with voxel- and BEV-based feature extraction, our architecture is faster and can handle different input resolutions by design. Experiments on SemanticKITTI and KITTI-360 show that our method achieves completion performance on par with the state of the art while completing a full scene in 0.1 seconds, which is 2.3 times faster than the fastest prior method. This matches the 10 Hz acquisition rate of typical automotive LiDAR sensors, taking a step toward real-time LiDAR scene completion.
Driving video generation models support autonomous-driving development by predicting controllable future scenes for simulation, planning evaluation, and offline data generation. Diffusion-based driving generators repeatedly evaluate large backbones across denoising steps, which limits generation throughput. Existing diffusion acceleration methods reduce this cost, but general-purpose designs omit driving signals available before generation, such as ego speed and planned trajectories. Experiments across driving motions show that cache tolerance varies with ego translation and rotation, denoising progress, and consecutive reuse length. We propose DriveCache, a training-free, action-aware controller that uses planned motion to allocate reuse across scenes and dynamic programming to place it across denoising steps under a calibrated response budget. A causal drift check refreshes features and replans the remaining schedule when generation departs from calibration. Across three generator configurations, DriveCache improves the overall fidelity-efficiency trade-off over evaluated cache methods. Our code will be publicly available.
4D occupancy forecasting models the spatio-temporal evolution of 3D scenes and is crucial for autonomous driving, especially for corner-case simulation. Existing methods often rely on discrete tokenization followed by autoregressive prediction, yet struggle with geometric distortion in static structures and inconsistent temporal coherence over the forecasting horizon. In this work, we propose a Geometry-Aware Spatio-Temporal context modeling method (GAST) for 4D occupancy forecasting, built upon progressive explicit-implicit generation and dual-path spatio-temporal modeling. Specifically, the generation module produces per-frame occupancy with high geometric fidelity and semantic plausibility through pose-driven warping, motion-aware feature modulation, and attention-based feature refinement. Subsequently, the spatio-temporal module enhances spatial consistency through global context aggregation while capturing scene evolution through temporal dynamics extraction. This unified design enables joint optimization of historical reconstruction and future forecasting in an end-to-end manner. Extensive experiments on Occ3D-nuScenes demonstrate the superiority of our method, outperforming the state-of-the-art by 7.67% in mIoU and 6.44% in IoU with a 2.84x speedup, while maintaining strong performance in long-term forecasting.
Although recent trajectory prediction and end-to-end autonomous driving methods improve robustness in urban environments, they still lack meaningful controllability. Existing benchmarks either provide no persona-conditioned annotations or support only a single urgency spectrum (i.e., emergency, normal, relaxed), which cannot distinguish personas that share the same urgency level but require different driving dynamics. To address this, we propose (i) the Persona-Conditioned Trajectory (PCT) dataset, which decomposes driving personas along two axes, Temporal Urgency and Ride Comfort, and combines three levels of each to form a grid of nine personas, each paired with natural-language descriptions and trajectories, and (ii) PersonaDrive, a framework that can learn driving personas from language and can generate persona-specific trajectories. PersonaDrive incorporates Persona-Conditioned Anchor Transform (PCAT), which hierarchically reshapes anchors along both axes, and Persona-Conditioned Multi-Modal Fusion (PCMF) for BEV-level persona fusion. Training is supervised by a Hierarchical Guide Loss enforcing axis-aligned physical orderings and an Axis-Decomposed Diversity Loss preventing diagonal mode collapse. Experimental results show that PersonaDrive consistently improves over the compared baselines across multi-dimensional scenarios. The code and PCT dataset are available at https://github.com/VisualAIKHU/PersonaDrive
Mohamed Abdelsamad, Bin Yang, Michael Ulrich +4cs.CV
3D object detection from LiDAR point clouds is a core problem in autonomous driving. Recent advances in self-supervised learning (SSL) enable scalable pretraining and transfers well to per-point tasks such as semantic and panoptic segmentation, but transfer to 3D detection remains weaker. We analyze recent SSL methods and find that most objectives are defined only on measured LiDAR returns from visible surfaces, leaving occluded and unobserved regions unconstrained. This visible-surface bias can be sufficient for point-wise prediction, but 3D detection requires robustness to missing structure. To address this gap, we propose GhostPoint, an SSL framework that hallucinates latent features in local neighborhoods around discovered instances, generated via a novel instance voxel dilation. In GhostPoint, an encoder processes observed returns, and an additional predictor infers neighborhood representations from observed context. In addition to standard encoder-level supervision, we introduce a predictor-level supervision scheme on sampled voxels from generated neighborhoods. Specifically, observed (visible/masked) voxels match teacher-encoder targets, while unobserved voxels match teacher-predictor hallucinations. This design encourages the learned representation to explicitly model structure beyond observed returns. Extensive evaluations on nuScenes and Waymo demonstrate that our method achieves state-of-the-art performance, consistently improving downstream 3D detection, especially under sparse scans and limited labels.
Perception is a fundamental component of autonomous driving systems. While LiDAR-based methods have achieved remarkable progress in object detection, their reliability can degrade under adverse weather conditions. Radar point clouds provide a robust alternative due to their resilience to bad weather and low-illumination scenarios. However, radar point clouds are typically sparse, unordered, and less informative than LiDAR data, making it challenging to directly apply existing LiDAR-based perception methods. To address these challenges, we propose IRGNN, an Invariant Radar Graph Neural Network for radar point cloud object detection. IRGNN first reconstructs radar point clouds into graph representations using translation- and rotation-invariant feature designs, enabling robust modeling of sparse radar measurements. It then employs an improved message passing neural network (MPNN) with residual connections and a virtual node layer to enhance local feature propagation and global context modeling. Finally, task-specific heads are applied to the learned graph representations for object classification and bounding box prediction. Experimental results on the RadarScenes dataset show that IRGNN outperforms existing radar-based object detection methods and achieves competitive performance. In addition, IRGNN significantly reduces computational cost and memory usage during inference, demonstrating its effectiveness and practical potential for efficient radar-based perception in autonomous driving.
Camera-based autonomous driving perception requires a shared representation that preserves metric 3D structure across synchronized multi-camera streams. However, existing image-based frameworks often rely on backbones pretrained for semantic recognition, and introduce 3D geometry through downstream task-specific modules. As a result, their shared representations may fail to preserve explicit metric geometry and consistent 3D scene structure. In this paper, we present a Geometry-grounded Unified 3D Perception (GeoUP) framework that adapts the reconstruction-oriented latent of VGGT to calibrated, streaming multi-camera driving scenes. GeoUP factorizes cross-image interaction into self, temporal, and view attention to capture structurally distinct temporal and cross-view correspondences. It further injects calibration-aware raymap encodings to provide metric scale and camera geometry. The resulting geometry-grounded latent is decoded for metric depth estimation, 3D object detection, and semantic occupancy prediction, corresponding to surface-, instance-, and volume-level readouts of the same 3D scene. Through joint multi-task and multi-dataset training, GeoUP effectively leverages heterogeneous annotations and generalizes across diverse sensor configurations and perception ranges. Extensive experiments on nuScenes, Argoverse 2, Waymo, KITTI, and DDAD demonstrate that GeoUP achieves SOTA performance across detection, occupancy, and depth estimation. These results validate the effectiveness of geometry-grounded representations for unified 3D driving perception.
Dense metric depth prediction from cameras and millimeter-wave radar offers a cost-effective sensing solution for autonomous systems. However, radar measurements are inherently sparse and susceptible to clutter, multipath reflections, and projection errors. While aggregating multiple radar frames provides denser metric cues, it also introduces temporal misalignment and dynamic-object interference. Directly propagating such unreliable measurements can therefore corrupt large regions of the predicted depth map. To address this issue, we propose RbFT-Net, an end-to-end rectify-before-fuse framework for multi-frame 4D radar-camera depth completion. Rather than assuming accumulated radar returns to be accurate, RbFT-Net treats them as noisy temporal anchor candidates. An image-conditioned rectification module jointly corrects their image-plane locations and metric depths while estimating pointwise reliability. The rectified anchors are then selectively propagated before high-level multi-modal fusion, suppressing the influence of unreliable measurements. Experiments on ZJU-4DRadarCam and a newly collected 4D radar-camera-LiDAR dataset show that RbFT-Net consistently outperforms the evaluated independent radar-camera methods and remains competitive with plug-in pipelines using auxiliary monocular depth models. Cross-platform evaluation and component analyses further support the effectiveness of the proposed rectification and reliability-aware propagation strategy.
Driving in the real world is open-world: a car may encounter a fallen mattress, a deer, or other objects outside its training data. Naming them is not enough. The system must know how to treat each region: can it drive over it, and how severe would a collision be? We therefore shift scene perception from category labels to dense action-relevant attributes, where each pixel is labeled by how it should affect motion rather than by object name. We instantiate this general formulation with two ordered attributes: 7-rank drivability and 5-rank vulnerability. We read Qwen3.5 image-token hidden states directly as a spatial semantic representation. A lightweight boundary-aware decoder then turns this coarse token grid into sharp full-resolution attribute maps. The whole process requires neither autoregressive text generation nor an external mask model such as SAM. We train on dense attribute labels built in CARLA and test transfer to real scenes and to novel obstacles never seen in training. We compare with vision-only segmenters trained on the same attributes and prompted VLM segmenters. Our model matches strong vision-only segmenters on familiar categories and improves transfer to real open-world anomalies, reaching 69.4% mean vulnerability-rank recall versus 57.1% for the best vision-only baseline and 53.9% for the best prompted VLM baseline. These results show that VLM image tokens provide useful semantic cues for transferring driving attributes to objects outside the training vocabulary.
Continuous and reliable localization is essential for autonomous driving. Cross-view visual localization matches ground images with satellite maps, providing complementary localization cues for pipelines that depend on Global Navigation Satellite System (GNSS) signals and high-definition (HD) maps. Most existing cross-view visual localization methods process each frame independently, leaving temporal information underused and limiting accuracy under dynamic occlusion, illumination variation, and repetitive textures. This study proposes a temporal-context-enhanced framework for cross-view sequence visual localization. The proposed recurrent cross-frame module aggregates historical context from the previous state to enhance the coarse ground feature of each current frame. These enhanced features facilitate satellite candidate-region classification, while hierarchical fine-grained features enable precise local offset estimation. On the CVIS dataset, the proposed method reduces mean localization error from 3.80 m to 1.57 m and increases R@1 m from 8.14% to 40.22%. Direct transfer to KITTI-CVL achieves a mean error of 2.61 m, with target-domain fine-tuning further reducing the mean error to 2.27 m. Zero-shot field experiments on a real-world vehicle achieve a mean error of 2.84 m and R@5 m of 96.86%. These results demonstrate that temporal context enhancement significantly improves cross-view localization accuracy and supports robust deployment on public benchmarks and real-world roads.
Emerging World-Action Models (WAMs) have demonstrated promising performance in autonomous driving by jointly modeling future driving scene evolution and trajectory planning. However, existing WAMs are typically trained with video data, which is only 2D projections of the underlying 4D driving scene. Consequently, WAMs fail to understand and capture the structure of 4D scenes and thus generate visually plausible yet 4D inconsistent future predictions that mislead downstream planning. To alleviate this issue, we present 4D-WAM, a model that leverages geometric foundation models for training-time supervision to enable 4D consistent world modeling. Specifically, we feed WAM-predicted future frames into a geometric foundation model, and use 4D-aware responses to define a 4D consistency loss. This loss encourages the model to understand, represent, and predict physically consistent 4D scenes during training, without additional inference cost. Moreover, we identify an early-decision phenomenon in WAMs and propose a decision-oriented timestep sampling strategy that emphasizes supervision at early, high-noise stages, where driving decisions are primarily formed. By propagating 4D supervision to this critical decision-formation phase, the proposed strategy further improves trajectory planning. Extensive experiments demonstrate that 4D-WAM effectively models 4D consistent scene evolution and achieves state-of-the-art performance on challenging NAVSIM-v1 and NAVSIM-v2 benchmarks.