Research map / Detection and segmentation / 3D detection and driving
GeoBEV: Learning Geometric BEV Representation for Multi-view 3D Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GeoBEV, a multi-view 3D detector that builds high-resolution BEV features via Radial-Cartesian sampling and supervises depth with an In-Box Label plus Centroid-Aware Inner Loss, reaching 66.2% NDS on nuScenes test.
Problem
Camera-only BEV detectors produce low-resolution, sparse BEV features and use LiDAR surface depth labels that fail to capture the true 3D geometry of objects, limiting detection accuracy.
Approach
Radial BEV features are formed by matrix multiplication of image features and depth scores, then bilinearly sampled into Cartesian BEV grids (RC-Sampling); depth is supervised by an In-Box Label (pseudo-points inside GT boxes are positive) with occlusion and background filtering, and a centroid-weighted focal loss (CAI Loss) emphasizes object interiors.
Key results
- On nuScenes test (VoVNet-99, 2-frame-free setting with 8 frames), GeoBEV reports 57.9% mAP and 66.2% NDS, the claimed state of the art.
- On nuScenes val with ResNet50 at 256x704 and 8 frames, GeoBEV reaches 47.9% mAP / 57.5% NDS, beating RayDN by 1.0% mAP and 1.2% NDS.
- Adding RC-Sampling and In-Box Label to BEVDepth raises mAP from 33.7% to 38.1% and NDS from 45.6% to 50.0% on the ablation setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GeoBEV instead of RayDN | nuScenes test, VoVNet-99 · nuScenes test | NDS: 0.645 -> 0.662 | Helped |
| Add RC-Sampling (ablation) | BEVDepth · nuScenes ablation | NDS: 0.456 -> 0.489 | Helped |
| Add In-Box Label and CAI Loss (ablation) | BEVDepth · nuScenes ablation | NDS: 0.456 -> 0.478 | Helped |
| Add RC-Sampling together with In-Box Label and CAI Loss (ablation) | BEVDepth · nuScenes ablation | NDS: 0.456 -> 0.500 | Helped |
| Use RC-Sampling and In-Box Label with the detector (ablation) | BEVDet · nuScenes ablation | NDS: 0.350 -> 0.391 | Helped |
| Use RC-Sampling and In-Box Label with the detector (ablation) | BEVStereo · nuScenes ablation | NDS: 0.474 -> 0.513 | Helped |
| Use DS 8 instead of DS 16 (ablation) | RC-Sampling, 256×256 BEV · nuScenes ablation | NDS: 0.482 -> 0.489 | Helped |
| Replace BEVPoolv2 with RC-Sampling (ablation) | BEVDepth, 256×256 BEV · nuScenes ablation | FPS: 16.6 -> 17.4 | Helped |
| Replace LiDAR depth labels with Vanilla In-Box Label (ablation) | BEVDepth · nuScenes ablation | mAP: 0.337 -> 0.345 | Helped |
| Add Sigmoid activation and Focal Loss (ablation) | Vanilla In-Box Label · nuScenes ablation | mAP: 0.345 -> 0.351 | Helped |
| Replace Focal Loss with CAI Loss (ablation) | In-Box Label with Sigmoid · nuScenes ablation | mAP: 0.356 -> 0.359 | Helped |
| Remove LiDAR supervision from background regions (ablation) | In-Box Label · nuScenes ablation | mAP: 0.351 -> 0.340 | Hurt |
Surprising findings
- Training with only In-Box labels and no LiDAR depth supervision still improves over no depth labels (mAP 33.0% to 35.1%), suggesting LiDAR points are not required for this supervision in the background-free setting.
- RC-Sampling matches Voxel-Sampling accuracy while using under 10% of its time and memory cost.
Limitations
- Relies on an external HTC instance-segmentation model pretrained on nuImages to filter background pseudo-points, adding dependency.
- Gains are reported on a BEVDepth-style pipeline and ablations use a shortened 24-epoch schedule without CBGS, so transfer to other settings is not fully established.
Experimental setup
Datasets: nuScenes nuImages
Benchmarks: nuScenes val nuScenes test
Metrics: mAP NDS mATE mASE mAOE mAVE mAAE FPS latency memory
Baselines: BEVDepth BEVDet BEVStereo BEVPoolv2 DFA3D RayDN StreamPETR BEVNext SA-BEV SOLOFusion FB-BEV OPEN
Builds on: BEVDepth LSS BEVPoolv2 Focal Loss BEV-Paste
Closest papers by meaning
- Towards Generalizable Multi-Camera 3D Object Detection via Perspective Rendering (AAAI 2025 · similarity 0.89)
- BEVNeXt: Reviving Dense BEV Frameworks for 3D Object Detection (CVPR 2024 · similarity 0.88)
- UniMODE: Unified Monocular 3D Object Detection (CVPR 2024 · similarity 0.88)
- SimDistill: Simulated Multi-Modal Distillation for BEV 3D Object Detection (AAAI 2024 · similarity 0.88)
- Instance-Aware Multi-Camera 3D Object Detection with Structural Priors Mining and Self-Boosting Learning (AAAI 2024 · similarity 0.88)
- CorrBEV: Multi-View 3D Object Detection by Correlation Learning with Multi-modal Prototypes (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.