Static adversarial camouflage exploits natural view-angle changes during relative motion to induce consistent feature drift in AV perception, leading to incorrect trajectory predictions and unnecessary braking.
hub Canonical reference
BEVDet: High-performance Multi-camera 3D Object Detection in Bird-Eye-View
Canonical reference. 71% of citing Pith papers cite this work as background.
39
Pith papers citing it
Background
71% of classified citations
abstract
Autonomous driving perceives its surroundings for decision making, which is one of the most complex scenarios in visual perception. The success of paradigm innovation in solving the 2D object detection task inspires us to seek an elegant, feasible, and scalable paradigm for fundamentally pushing the performance boundary in this area. To this end, we contribute the BEVDet paradigm in this paper. BEVDet performs 3D object detection in Bird-Eye-View (BEV), where most target values are defined and route planning can be handily performed. We merely reuse existing modules to build its framework but substantially develop its performance by constructing an exclusive data augmentation strategy and upgrading the Non-Maximum Suppression strategy. In the experiment, BEVDet offers an excellent trade-off between accuracy and time-efficiency. As a fast version, BEVDet-Tiny scores 31.2% mAP and 39.2% NDS on the nuScenes val set. It is comparable with FCOS3D, but requires just 11% computational budget of 215.3 GFLOPs and runs 9.2 times faster at 15.6 FPS. Another high-precision version dubbed BEVDet-Base scores 39.3% mAP and 47.2% NDS, significantly exceeding all published results. With a comparable inference speed, it surpasses FCOS3D by a large margin of +9.8% mAP and +10.0% NDS. The source code is publicly available for further research at https://github.com/HuangJunJie2017/BEVDet .
hub tools
citation-role summary
background 5
baseline 1
method 1
citation-polarity summary
representative citing papers
Introduces Plentiful CARLA Camera Rigs benchmark and two calibration-derived metrics (Rig Variance, Rig Contrastive Distance) showing geometric rig differences correlate with cross-rig performance drops in multi-view automotive perception.
DAPETR adds two learned adaptive modules to PETR for superior fisheye BEV detection on converted KITTI-360 data, outperforming PolarPETR while revealing negative interaction when both are combined.
Introduces a synchronized cross-view urban traffic dataset with drone ground truth for identity matching and monocular BEV localization tasks.
Bench2Drive-Robust is a new closed-loop benchmark that evaluates end-to-end autonomous driving models under deployment perturbations from camera failures, ego-state errors, and compute delays, showing substantial performance degradation beyond image-level tests.
EgoEV-HandPose uses stereo event cameras and a bird's-eye-view fusion module to achieve 30.54 mm MPJPE and 86.87% gesture accuracy on a new large-scale egocentric dataset, outperforming prior RGB and event methods especially in low light and occlusion.
PointForward uses sparse world-space 3D queries and scene graphs to deliver consistent single-pass reconstruction of dynamic driving scenes via point-aligned representations.
The paper organizes perception attacks on AVs into a new taxonomy, identifies gaps in fusion-aware defenses, and validates one cross-sensor vulnerability with a proof-of-concept simulation.
Dynamic token selection and training only 1.6 million parameters instead of over 300 million reduces computation by 48-55% and improves accuracy over prior state-of-the-art on the NuScenes dataset.
DinoRADE reports a radar-centered multi-class detection pipeline that fuses dense radar tensors with DINOv3 features via deformable attention and outperforms prior radar-camera methods by 12.1% on the K-Radar dataset across weather conditions.
TopoMaskV3 adds dense offset and height heads to produce standalone 3D road centerlines from masks and reports 28.5 OLS on a new geographically disjoint long-range benchmark.
TRIG factorizes multi-camera poses into ego-trajectory and static rig geometry, with decoupled supervision and sparse temporal-spatial attention, claiming SOTA metric depth, pose, and 3D reconstruction on five driving benchmarks.
DRVR uses range-view and geometry-aware voxel-view encoders plus fusion to deliver 5.4% higher mIoU and 2.1x faster inference than multi-sweep baselines on nuScenes-Occupancy from single sweeps.
UniTeD unifies perception and planning in autonomous driving via shared temporal diffusion with TTM and ARS modules, reporting SOTA results on benchmarks.
OVBEVSeg produces open-vocabulary bird's-eye-view semantic maps on nuScenes by projecting CLIP labels through 3D detections, constraining Gaussian splats with BEV occupancy, and distilling the geometry into a real-time student (15.3 mIoU novel, 0% novel GT).
WalkOCC bootstraps pseudo 3D occupancy labels from paired LiDAR-RGB sequences and jointly trains on unpaired monocular images for sidewalk robots, plus introduces the Sidewalk3D dataset.
DeGO decouples rigid and nonrigid motion in Gaussian occupancy prediction via factorized 4D distillation from VGGT, reporting SOTA results on Occ3D-NuScenes with 13.5% gains on human-centric cases.
HiPR improves 3D occupancy prediction by reparameterizing image-to-voxel projections using LiDAR-derived height priors to adapt sampling ranges to scene sparsity and height variations.
SimPB++ unifies multi-view 2D perspective and 3D BEV object detection in one model via an interactive hybrid decoder, reporting state-of-the-art results on nuScenes and long-range detection up to 150 m on Argoverse2.
OneDrive unifies heterogeneous decoding in a single VLM transformer decoder for end-to-end driving, achieving 0.28 L2 error and 0.18 collision rate on nuScenes plus 86.8 PDMS on NAVSIM.
CAM3DNet outperforms prior camera-based 3D detectors on nuScenes, Waymo and Argoverse by using three new modules to better mine multi-scale spatiotemporal features from 2D queries and pyramid maps.
ESCAPE combines spatio-temporal fusion mapping for depth-free 3D memory with a memory-driven grounding module and adaptive execution policy to reach 65.09% success on ALFRED test-seen long-horizon mobile manipulation tasks.
DVGT-2 is a streaming vision-geometry-action model that jointly reconstructs dense 3D geometry and plans trajectories online, achieving better reconstruction than prior batch methods while transferring directly to planning benchmarks without fine-tuning.
R4Det fuses 4D radar and camera inputs via panoramic depth fusion, deformable gated temporal fusion without ego pose, and instance-guided refinement to reach state-of-the-art 3D detection on TJ4DRadSet and VoD.
citing papers explorer
-
Still Camouflage, Moving Illusion: View-Induced Trajectory Manipulation in Autonomous Driving
Static adversarial camouflage exploits natural view-angle changes during relative motion to induce consistent feature drift in AV perception, leading to incorrect trajectory predictions and unnecessary braking.
-
Understanding Cross-Rig Generalization in Automotive Perception: a Multi-Rig Benchmark and Rig Variation Metrics
Introduces Plentiful CARLA Camera Rigs benchmark and two calibration-derived metrics (Rig Variance, Rig Contrastive Distance) showing geometric rig differences correlate with cross-rig performance drops in multi-view automotive perception.
-
Distortion-Aware PETR for BEV Object Detection with Mixed Pinhole-Fisheye Cameras
DAPETR adds two learned adaptive modules to PETR for superior fisheye BEV detection on converted KITTI-360 data, outperforming PolarPETR while revealing negative interaction when both are combined.
-
Cross-View Urban Traffic Dataset: Drone-Supervised Ground Truth for Monocular Bird's-Eye View Localization
Introduces a synchronized cross-view urban traffic dataset with drone ground truth for identity matching and monocular BEV localization tasks.
-
Bench2Drive-Robust: Benchmarking Closed-Loop Autonomous Driving under Deployment Perturbations
Bench2Drive-Robust is a new closed-loop benchmark that evaluates end-to-end autonomous driving models under deployment perturbations from camera failures, ego-state errors, and compute delays, showing substantial performance degradation beyond image-level tests.
-
EgoEV-HandPose: Egocentric 3D Hand Pose Estimation and Gesture Recognition with Stereo Event Cameras
EgoEV-HandPose uses stereo event cameras and a bird's-eye-view fusion module to achieve 30.54 mm MPJPE and 86.87% gesture accuracy on a new large-scale egocentric dataset, outperforming prior RGB and event methods especially in low light and occlusion.
-
PointForward: Feedforward Driving Reconstruction through Point-Aligned Representations
PointForward uses sparse world-space 3D queries and scene graphs to deliver consistent single-pass reconstruction of dynamic driving scenes via point-aligned representations.
-
SoK: The Next Frontier in AV Security: Systematizing Perception Attacks and the Emerging Threat of Multi-Sensor Fusion
The paper organizes perception attacks on AVs into a new taxonomy, identifies gaps in fusion-aware defenses, and validates one cross-sensor vulnerability with a proof-of-concept simulation.
-
Efficient Multi-View 3D Object Detection by Dynamic Token Selection and Fine-Tuning
Dynamic token selection and training only 1.6 million parameters instead of over 300 million reduces computation by 48-55% and improves accuracy over prior state-of-the-art on the NuScenes dataset.
-
DinoRADE: Full Spectral Radar-Camera Fusion with Vision Foundation Model Features for Multi-class Object Detection in Adverse Weather
DinoRADE reports a radar-centered multi-class detection pipeline that fuses dense radar tensors with DINOv3 features via deformable attention and outperforms prior radar-camera methods by 12.1% on the K-Radar dataset across weather conditions.
-
TopoMaskV3: 3D Mask Head with Dense Offset and Height Predictions for Road Topology Understanding
TopoMaskV3 adds dense offset and height heads to produce standalone 3D road centerlines from masks and reports 28.5 OLS on a new geographically disjoint long-range benchmark.
-
TRIG: Trajectory-Rig Decoupled Metric Geometry Learning
TRIG factorizes multi-camera poses into ego-trajectory and static rig geometry, with decoupled supervision and sparse temporal-spatial attention, claiming SOTA metric depth, pose, and 3D reconstruction on five driving benchmarks.
-
Semantic Occupancy Prediction with Dual Range-Voxel Representation
DRVR uses range-view and geometry-aware voxel-view encoders plus fusion to deliver 5.4% higher mIoU and 2.1x faster inference than multi-sweep baselines on nuScenes-Occupancy from single sweeps.
-
UniTeD: Unified Temporal Diffusion for Joint Perception and Planning in Autonomous Driving
UniTeD unifies perception and planning in autonomous driving via shared temporal diffusion with TTM and ARS modules, reporting SOTA results on benchmarks.
-
Open-Vocabulary BEV Segmentation with 3D-Aware Geometric Constraints
OVBEVSeg produces open-vocabulary bird's-eye-view semantic maps on nuScenes by projecting CLIP labels through 3D detections, constraining Gaussian splats with BEV occupancy, and distilling the geometry into a real-time student (15.3 mIoU novel, 0% novel GT).
-
Monocular 3D Occupancy Perception for Robots on Sidewalks via Hybrid 2D-3D Learning
WalkOCC bootstraps pseudo 3D occupancy labels from paired LiDAR-RGB sequences and jointly trains on unpaired monocular images for sidewalk robots, plus introduces the Sidewalk3D dataset.
-
Deformable Gaussian Occupancy: Decoupling Rigid and Nonrigid Motion with Factorized Distillation
DeGO decouples rigid and nonrigid motion in Gaussian occupancy prediction via factorized 4D distillation from VGGT, reporting SOTA results on Occ3D-NuScenes with 13.5% gains on human-centric cases.
-
Height-Guided Projection Reparameterization for Camera-LiDAR Occupancy
HiPR improves 3D occupancy prediction by reparameterizing image-to-voxel projections using LiDAR-derived height priors to adapt sampling ranges to scene sparsity and height variations.
-
SimPB++: Simultaneously Detecting 2D and 3D Objects from Multiple Cameras
SimPB++ unifies multi-view 2D perspective and 3D BEV object detection in one model via an interactive hybrid decoder, reporting state-of-the-art results on nuScenes and long-range detection up to 150 m on Argoverse2.
-
OneDrive: Unified Multi-Paradigm Driving with Vision-Language-Action Models
OneDrive unifies heterogeneous decoding in a single VLM transformer decoder for end-to-end driving, achieving 0.28 L2 error and 0.18 collision rate on nuScenes plus 86.8 PDMS on NAVSIM.
-
CAM3DNet: Comprehensively mining the multi-scale features for 3D Object Detection with Multi-View Cameras
CAM3DNet outperforms prior camera-based 3D detectors on nuScenes, Waymo and Argoverse by using three new modules to better mine multi-scale spatiotemporal features from 2D queries and pyramid maps.
-
ESCAPE: Episodic Spatial Memory and Adaptive Execution Policy for Long-Horizon Mobile Manipulation
ESCAPE combines spatio-temporal fusion mapping for depth-free 3D memory with a memory-driven grounding module and adaptive execution policy to reach 65.09% success on ALFRED test-seen long-horizon mobile manipulation tasks.
-
DVGT-2: Vision-Geometry-Action Model for Autonomous Driving at Scale
DVGT-2 is a streaming vision-geometry-action model that jointly reconstructs dense 3D geometry and plans trajectories online, achieving better reconstruction than prior batch methods while transferring directly to planning benchmarks without fine-tuning.
-
R4Det: 4D Radar-Camera Fusion for High-Performance 3D Object Detection
R4Det fuses 4D radar and camera inputs via panoramic depth fusion, deformable gated temporal fusion without ego pose, and instance-guided refinement to reach state-of-the-art 3D detection on TJ4DRadSet and VoD.
-
TFusionOcc: T-Primitive Based Object-Centric Multi-Sensor Fusion Framework for 3D Occupancy Prediction
TFusionOcc uses a family of Student's t-distribution T-primitives and a T-mixture model for multi-sensor 3D occupancy prediction, reporting state-of-the-art results on nuScenes.
-
DriveLaW:Unifying Planning and Video Generation in a Latent Driving World
DriveLaW unifies video world modeling and trajectory planning by injecting video-generator latents into a diffusion planner, achieving SOTA video prediction and a new record on the NAVSIM planning benchmark.
-
DriveVLA-W0: World Models Amplify Data Scaling Law in Autonomous Driving
DriveVLA-W0 adds world modeling to predict future images in VLA models, overcoming sparse action supervision and amplifying data scaling laws on NAVSIM benchmarks and a large in-house dataset.
-
BePo: Dual Representation for 3D Occupancy Prediction
BePo proposes a dual BEV and sparse-points representation with cross-attention fusion for more accurate and efficient 3D occupancy prediction on autonomous driving benchmarks.
-
RQR3D: Reparametrizing the regression targets for BEV-based 3D object detection
RQR3D reparametrizes oriented bounding box regression in BEV 3D detection as regressing a horizontal box plus corner offsets and achieves SOTA camera-radar performance on nuScenes with 67.5 NDS and 59.7 mAP.
-
CascadeOcc: Rethinking 3D Occupancy World Models with Cascaded VQ Representations
CascadeOcc uses cascaded VQ representations in an autoregressive framework with a TimeMixer for multi-scale spatial and temporal modeling, achieving top results among vision-centric methods on 4D occupancy and planning benchmarks.
-
VGGT-Occ: Geometry-Grounded and Density-Aware Gated Fusion for 3D Occupancy Prediction
VGGT-Occ embeds geometric tokens via PA-DA and uses sequential coarse-to-fine gated fusion to reach 33.00% IoU and 21.08% mIoU on SurroundOcc-nuScenes while using only ~41M parameters in the occupancy head.
-
InterFuserDVS: Event-Enhanced Sensor Fusion for Safe RL-Based Decision Making
Integrating DVS event data into InterFuser through token fusion yields a driving score of 77.2 and 100% route completion on CARLA benchmarks, indicating improved robustness in dynamic conditions.
-
SemLT3D: Semantic-Guided Expert Distillation for Camera-only Long-Tailed 3D Object Detection
SemLT3D introduces semantic-guided expert distillation with a language MoE module and CLIP projection to enrich features for long-tailed classes in camera-only 3D detection.
-
Radar-Camera BEV Multi-Task Learning with Cross-Task Attention Bridge for Joint 3D Detection and Segmentation
CTAB bidirectional deformable attention between detection and segmentation BEV branches improves 7-class mIoU by 0.6 at neutral detection on nuScenes radar-camera multi-task learning.
-
Not All Agents Matter: From Global Attention Dilution to Risk-Prioritized Game Planning
GameAD models autonomous driving as a risk-prioritized game among agents via Risk-Aware Topology Anchoring, Minimax Risk-Aware Sparse Attention and related components, yielding safer trajectories than prior end-to-end methods on nuScenes and Bench2Drive.
-
Multi-Modal Sensor Fusion using Hybrid Attention for Autonomous Driving
MMF-BEV fuses camera and radar branches with deformable self- and cross-attention, outperforming unimodal baselines on the VoD 4D radar dataset through a two-stage training process.
-
BEVPredFormer: Spatio-temporal Attention for BEV Instance Prediction in Autonomous Driving
BEVPredFormer uses attention-based temporal processing and 3D camera projection to match or exceed prior methods on nuScenes for BEV instance prediction.
-
Fast-BEV++: Fast by Algorithm, Deployable by Design
Fast-BEV++ achieves at least 3x speedup over Fast-BEV, a new SOTA of 0.488 NDS on nuScenes 3D detection, and over 134 FPS inference by redesigning the core transformation pipeline and adding a learnable depth module.
-
U-ViLAR: Uncertainty-Aware Visual Localization for Autonomous Driving via Differentiable Association and Registration
U-ViLAR achieves robust visual localization for autonomous driving by mapping features to BEV and applying perceptual uncertainty-guided association plus localization uncertainty-guided registration.