Research map / 3D vision / 3D generation
EvOcc: Accurate Semantic Occupancy for Automated Driving Using Evidence Theory
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 EvOcc, an evidence-theory (Dempster-Shafer) framework that builds semantic occupancy ground truth from noisy, partially labeled LiDAR and trains camera-based occupancy models with a BBA cross-entropy loss.
Problem
Ground-truth semantic occupancy maps built from LiDAR with heuristics are inaccurate and Bayesian occupancy models cannot represent uncertainty from occlusions or conflicting measurements, limiting camera-based semantic occupancy prediction for automated driving.
Approach
Extends evidential (Dempster-Shafer) occupancy mapping from binary occupancy to semantic classes via a new per-voxel basic belief assignment built from class-specific spherical reflection and transmission counts, and supervises an off-the-shelf image-based occupancy network directly with ground-truth BBAs using a BBA cross-entropy loss.
Key results
- Ground-truth occupancy at 0.4 m voxels achieves depth MAE 1.21 m vs 1.89 m for Occ3D (EvOcc vs Occ3D on nuScenes).
- Semantic ray-based mIoU of ground truth at 0.4 m is 67.0% vs 51.2% for Occ3D, and 74.4% vs 54.1% for OpenOccupancy at 0.2 m.
- BEVDet4D-Occ trained with EvOcc supervision improves predicted mIoU from 43.1% to 48.6% and depth MAE from 2.33 m to 1.66 m versus Occ3D supervision.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace heuristic occupancy GT generation with EvOcc evidential semantic mapping | nuScenes semantic occupancy GT, voxel size 0.4 m · nuScenes | ray-based mIoU: 51.2 -> 67.0 | Helped |
| Replace heuristic occupancy GT generation with EvOcc evidential semantic mapping | nuScenes semantic occupancy GT, voxel size 0.2 m · nuScenes | ray-based mIoU: 54.1 -> 74.4 | Helped |
| Train BEVDet4D-Occ with EvOcc supervision instead of Occ3D supervision | BEVDet4D-Occ with ResNet-50 · nuScenes semantic occupancy prediction, voxel size 0.4 m | ray-based mIoU: 43.1 -> 48.6 | Helped |
| Replace Kälble et al.'s occupancy GT mapping with EvOcc | nuScenes semantic occupancy GT, voxel size 0.4 m · nuScenes | MAE: 1.25 -> 1.21 | Helped |
| Replace Kälble et al.'s occupancy GT mapping with EvOcc | nuScenes semantic occupancy GT, voxel size 0.2 m · nuScenes | MAE: 0.91 -> 0.84 | Helped |
| Train BEVDet4D-Occ with EvOcc supervision instead of Occ3D supervision | BEVDet4D-Occ with ResNet-50 · nuScenes semantic occupancy prediction, voxel size 0.4 m | MAE: 2.33 -> 1.66 | Helped |
Limitations
- Evaluation relies on LiDAR measurements as proxy reference since no perfect ground-truth occupancy exists.
- Semantic labels in nuScenes are available for only one in ten LiDAR scans, so measurements are partially annotated.
Experimental setup
Datasets: nuScenes Occ3D Waymo (supplementary)
Benchmarks: nuScenes
Metrics: MAE RMSE delta threshold accuracy ray-based IoU mIoU class-and-depth accuracy
Baselines: Occ3D SurroundOcc Scene as Occupancy OpenOccupancy Kälble et al. 2024
Builds on: Kälble et al. evidential occupancy mapping BEVDet4D-Occ Dempster-Shafer theory Dezert-Dambreville BBA cross-entropy
Closest papers by meaning
- Accurate Training Data for Occupancy Map Prediction in Automated Driving Using Evidence Theory (CVPR 2024 · similarity 0.93)
- Achieving Speed-Accuracy Balance in Vision-based 3D Occupancy Prediction via Geometric-Semantic Disentanglement (AAAI 2025 · similarity 0.88)
- CSV-Occ: Fusing Multi-frame Alignment for Occupancy Prediction with Temporal Cross State Space Model and Central Voting Mechanism (ICML 2025 · similarity 0.88)
- Collaborative Semantic Occupancy Prediction with Hybrid Feature Fusion in Connected Automated Vehicles (CVPR 2024 · similarity 0.88)
- SDGOCC: Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction (CVPR 2025 · similarity 0.87)
- OctOcc: High-Resolution 3D Occupancy Prediction with Octree (AAAI 2024 · 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.