Research map / 3D vision / Depth and stereo
Accurate Training Data for Occupancy Map Prediction in 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 an evidence-theory (Dempster-Shafer) method that converts LiDAR scans into more accurate 3D occupancy ground truth with uncertainty, and uses it with a loss weighting to train a camera-based occupancy predictor.
Problem
Occupancy ground-truth data for camera-based automated driving occupancy prediction, generated from LiDAR scans by existing pipelines, is of low quality; visibility masks underestimate visibility and lack uncertainty.
Approach
Maps LiDAR reflections and transmissions into spherical grids, aggregates them multi-frame into a Cartesian grid, and assigns evidence masses (occupied, free, uncertain) via evidence theory with sensor-specific false positive/negative probabilities; the uncertainty mass is used as a loss weight for training occupancy predictors.
Key results
- On nuScenes, the proposed grid at 0.4 m reduces depth MAE from 1.80 m (Occ3D) to 1.25 m; at 0.2 m it reaches 0.91 m vs 1.89 m for OpenOccupancy.
- On Waymo at 0.4 m, MAE drops from 3.23 m (Occ3D) to 1.51 m (about 53% improvement).
- Fine-tuning BEVDet4D-Occ-R50 on the new occupancy GT with the uncertainty-weighted CE loss lowers MAE from 1.89 m to 1.42 m (about 25%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace Scene as Occupancy ground truth with evidence-theory occupancy mapping | Occupancy ground-truth generation · nuScenes, 0.5 m baseline vs 0.4 m ours | MAE: 2.44 -> 1.25 | Helped |
| Replace SurroundOcc ground truth with evidence-theory occupancy mapping | Occupancy ground-truth generation · nuScenes, 0.5 m baseline vs 0.4 m ours | MAE: 2.36 -> 1.25 | Helped |
| Replace Occ3D ground truth with evidence-theory occupancy mapping | Occupancy ground-truth generation · nuScenes, 0.4 m voxels | MAE: 1.80 -> 1.25 | Helped |
| Replace OpenOccupancy ground truth with evidence-theory occupancy mapping | Occupancy ground-truth generation · nuScenes, 0.2 m voxels | MAE: 1.89 -> 0.91 | Helped |
| Replace Occ3D ground truth with evidence-theory occupancy mapping | Occupancy ground-truth generation · Waymo, 0.4 m voxels | MAE: 3.23 -> 1.51 | Helped |
| Train BEVDet4D R50 with our occupancy GT and uncertainty-weighted BCE instead of Occ3D GT | BEVDet4D R50 · nuScenes validation | MAE: 1.89 -> 1.44 | Helped |
| Train BEVDet4D R50 with our occupancy GT and uncertainty-weighted CE instead of Occ3D GT | BEVDet4D R50 · nuScenes validation | MAE: 1.89 -> 1.42 | Helped |
| Fine-tune on Occ3D occupancy GT (ablation) | BEVDet4D R50 · nuScenes validation | MAE: 1.93 -> 1.89 | Helped |
| Use three-class occupied/free/uncertain CE supervision instead of binary BCE supervision (ablation) | BEVDet4D R50 trained with our occupancy GT · nuScenes validation | MAE: 1.44 -> 1.42 | Helped |
Surprising findings
- Fine-tuning with Occ3D's occupancy maps yields only small improvements over the baseline, while the new GT gives significant gains.
- Existing visibility masks substantially underestimate visibility, preventing supervision in critical regions.
Limitations
- Semantic information is not incorporated into the preprocessing and training pipeline.
- Hyperparameters (p_FN, p_FP) are chosen heuristically via grid search on a small nuScenes subset.
Experimental setup
Datasets: nuScenes Waymo Occ3D occupancy GT
Benchmarks: nuScenes Waymo
Metrics: MAE RMSE RMSE_log delta < 1.25 accuracy
Baselines: Scene as Occupancy SurroundOcc Occ3D OpenOccupancy BEVDet4D-Occ-R50 baseline
Builds on: Occ3D BEVDet4D evidential semantic grid mapping (Richter et al.) Dempster-Shafer theory
Closest papers by meaning
- EvOcc: Accurate Semantic Occupancy for Automated Driving Using Evidence Theory (CVPR 2025 · similarity 0.93)
- SDGOCC: Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction (CVPR 2025 · similarity 0.87)
- UnO: Unsupervised Occupancy Fields for Perception and Forecasting (CVPR 2024 · similarity 0.87)
- OccProphet: Pushing the Efficiency Frontier of Camera-Only 4D Occupancy Forecasting with an Observer-Forecaster-Refiner Framework (ICLR 2025 · similarity 0.87)
- UniScene: Unified Occupancy-centric Driving Scene Generation (CVPR 2025 · similarity 0.87)
- Achieving Speed-Accuracy Balance in Vision-based 3D Occupancy Prediction via Geometric-Semantic Disentanglement (AAAI 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.