Research map / Detection and segmentation / 3D detection and driving
BEVSpread: Spread Voxel Pooling for Bird's-Eye-View Representation in Vision-based Roadside 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 BEVSpread, a plug-in voxel pooling that spreads each frustum point's image features to its top-k neighboring BEV grids with Gaussian distance/depth-dependent weights, improving roadside monocular 3D detection on DAIR-V2X-I and Rope3D.
Problem
Position approximation error in frustum-based voxel pooling for vision-based roadside 3D object detection, where each frustum point's features are collapsed onto a single BEV grid center, which is irrecoverable and worsened by sparse BEV grids.
Approach
Instead of accumulating a frustum point's features into one BEV grid, spread them to the k nearest BEV grid centers with Gaussian weights whose variance scales with predicted depth, so distance decay is slower for distant objects; implemented with customized CUDA kernels for comparable latency.
Key results
- On DAIR-V2X-I validation, BEVSpread (on BEVHeight, ResNet-101) improves over BEVHeight by +1.29/+5.32/+2.41 AP (vehicle easy / pedestrian easy / cyclist easy) at 0-100m.
- Proof experiment: spread voxel pooling recovers point position with 0.003 MSE vs 0.095 for original voxel pooling when neighbors k>=3.
- On nuScenes val, BEVDepth reproduction + spread improves NDS from 0.432 to 0.450.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add spread voxel pooling to BEVHeight | BEVHeight · DAIR-V2X-I, vehicle (Middle) | AP3D|R40: 65.77 -> 66.82 | Helped |
| Add spread voxel pooling to BEVHeight | BEVHeight · DAIR-V2X-I, pedestrian (Middle) | AP3D|R40: 39.29 -> 44.51 | Helped |
| Add spread voxel pooling to BEVHeight | BEVHeight · DAIR-V2X-I, cyclist (Middle) | AP3D|R40: 60.08 -> 63.50 | Helped |
| Add spread voxel pooling to BEVDepth (ablation) | BEVDepth · DAIR-V2X-I, vehicle (Middle) | AP3D|R40: 60.37 -> 64.09 | Helped |
| Add spread voxel pooling to BEVDepth (ablation) | BEVDepth · DAIR-V2X-I, pedestrian (Middle) | AP3D|R40: 20.84 -> 29.27 | Helped |
| Add spread voxel pooling to BEVDepth (ablation) | BEVDepth · DAIR-V2X-I, cyclist (Middle) | AP3D|R40: 40.34 -> 48.53 | Helped |
| Add spread voxel pooling to BEVHeight (ablation) | BEVHeight (ResNet-50) · DAIR-V2X-I, vehicle (Middle) | AP3D|R40: 64.54 -> 65.80 | Helped |
| Add spread voxel pooling to BEVHeight (ablation) | BEVHeight (ResNet-50) · DAIR-V2X-I, pedestrian (Middle) | AP3D|R40: 25.79 -> 31.00 | Helped |
| Add spread voxel pooling to BEVHeight (ablation) | BEVHeight (ResNet-50) · DAIR-V2X-I, cyclist (Middle) | AP3D|R40: 48.21 -> 56.34 | Helped |
| Include depth in the spread weight function (ablation) | BEVSpread (ResNet-50) · DAIR-V2X-I, pedestrian (Middle) | AP3D|R40: 29.94 -> 30.97 | Helped |
| Include learnable α in the spread weight function (ablation) | BEVSpread (ResNet-50) · DAIR-V2X-I, cyclist (Middle) | AP3D|R40: 54.65 -> 55.64 | Helped |
| Increase neighbor count from k=1 to k=2 (ablation) | BEVSpread (ResNet-101) · DAIR-V2X-I | Avg-AP: 56.69 -> 58.68 | Helped |
Surprising findings
- Spread voxel pooling with k=2 adds essentially no latency overhead on ResNet-101 (73.9 ms vs 74.3 ms baseline) thanks to CUDA optimization.
Limitations
- Increases latency as neighbor count k grows (k=6 gives +15.2% total latency).
- Neighbor coordinates are computed online; deployment would need a precomputed look-up table, which is not implemented.
- Evaluated mainly on roadside datasets; nuScenes gains are on a reproduced baseline with modest changes.
Experimental setup
Datasets: DAIR-V2X-I Rope3D nuScenes
Benchmarks: DAIR-V2X-I Rope3D nuScenes
Metrics: AP3D|R40 NDS mAP Avg-AP latency
Baselines: BEVHeight BEVDepth BEVFormer ImVoxelNet PointPillars SECOND MVXNet
Builds on: BEVHeight BEVDepth LSS
Closest papers by meaning
- BEVHeight: A Robust Framework for Vision-Based Roadside 3D Object Detection (CVPR 2023 · similarity 0.91)
- Every View Counts: Cross-View Consistency in 3D Object Detection with Hybrid-Cylindrical-Spherical Voxelization (NeurIPS 2020 · similarity 0.88)
- BEV-SAN: Accurate BEV 3D Object Detection via Slice Attention Networks (CVPR 2023 · similarity 0.88)
- Rope3D: The Roadside Perception Dataset for Autonomous Driving and Monocular 3D Object Detection Task (CVPR 2022 · similarity 0.88)
- SOGDet: Semantic-Occupancy Guided Multi-View 3D Object Detection (AAAI 2024 · similarity 0.87)
- 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.