Research map / 3D vision / Gaussian splatting
ODG: Occupancy Prediction Using Dual Gaussians
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 ODG, a 3D occupancy predictor using dual static/dynamic Gaussian queries with hierarchical coarse-to-fine Gaussian refinement and rendering-based depth/semantic supervision, reaching SOTA on Occ3D-nuScenes and Occ3D-Waymo.
Problem
Camera-based 3D occupancy prediction is hard to scale with dense grids, and single-set sparse query methods handle heterogeneous static and dynamic scene objects poorly.
Approach
Two sparse Gaussian query sets (static and dynamic), with dynamic queries extended by box and velocity attributes and motion-compensated sampling; queries are decoded coarse-to-fine into many Gaussians, joined by self-attention over concatenated queries, and supervised by rendered LiDAR-projected depth and semantic maps via 3D Gaussian splatting.
Key results
- ODG-L (8f) reaches 38.18 mIoU and 42.3 RayIoU on Occ3D-nuScenes validation, versus 36.2 mIoU for OPUS-L (8f).
- ODG-T (8f) reaches 35.54 mIoU and 39.2 RayIoU at 20.1 FPS on Occ3D-nuScenes.
- ODG-L reaches 21.35 mIoU and 25.9 RayIoU on Occ3D-Waymo (trained on 20% of data), +2.35 mIoU over OPUS-L.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ODG-L instead of OPUS-L | Occ3D-nuScenes occupancy prediction · Occ3D-nuScenes validation | mIoU: 36.2 -> 38.18 | Helped |
| use ODG-L instead of OPUS-L | Occ3D-nuScenes occupancy prediction · Occ3D-nuScenes validation | RayIoU: 41.2 -> 42.3 | Helped |
| use ODG-L instead of OPUS-L | Occ3D-Waymo occupancy prediction · Occ3D-Waymo validation | mIoU: 19.00 -> 21.35 | Helped |
| use ODG-L instead of OPUS-L | Occ3D-Waymo occupancy prediction · Occ3D-Waymo validation | RayIoU: 24.7 -> 25.9 | Helped |
| add motion compensation (ablation) | ODG-T · Occ3D-nuScenes | mIoU: 30.80 -> 31.78 | Helped |
| add motion compensation (ablation) | ODG-T · Occ3D-nuScenes | RayIoU₁ₘ: 27.9 -> 28.7 | Helped |
| add query attention after motion compensation (ablation) | ODG-T · Occ3D-nuScenes | mIoU: 31.78 -> 32.13 | Helped |
| add rendering supervision after motion compensation and query attention (ablation) | ODG-T · Occ3D-nuScenes | mIoU: 32.13 -> 32.82 | Helped |
| add rendering supervision after motion compensation and query attention (ablation) | ODG-T · Occ3D-nuScenes | RayIoU₂ₘ: 37.8 -> 40.5 | Helped |
| use self-attention on concatenated dynamic and static query features instead of cross attention (ablation) | ODG-T · Occ3D-nuScenes | mIoU: 31.95 -> 32.13 | Helped |
| use self-attention on concatenated dynamic and static query features instead of cross attention (ablation) | ODG-T · Occ3D-nuScenes | RayIoU: 36.3 -> 36.7 | Helped |
| add dynamic-object motion compensation to ego-motion compensation (ablation) | ODG-T · Occ3D-nuScenes | mIoU: 31.17 -> 31.78 | Helped |
Surprising findings
- Ego-motion-only compensation for dynamic Gaussians lowers mIoU by 0.71 relative to adding object-motion compensation.
Limitations
- Gaussian scale, rotation and opacity are optimized only through rendering loss, not occupancy ground truth.
- Aggregating Gaussians into occupancy for direct supervision is costly.
Experimental setup
Baselines: RenderOcc GaussRender GaussTR SparseOcc OPUS GaussianFlowOcc CTF-Occ BEVFormer
Builds on: 3D Gaussian Splatting GaussTR SparseOcc DETR
Closest papers by meaning
- GaussianFormer-2: Probabilistic Gaussian Superposition for Efficient 3D Occupancy Prediction (CVPR 2025 · similarity 0.92)
- GaussianWorld: Gaussian World Model for Streaming 3D Occupancy Prediction (CVPR 2025 · similarity 0.92)
- VoxelSplat: Dynamic Gaussian Splatting as an Effective Loss for Occupancy and Flow Prediction (CVPR 2025 · similarity 0.91)
- OctOcc: High-Resolution 3D Occupancy Prediction with Octree (AAAI 2024 · similarity 0.90)
- QuadricFormer: Scene as Superquadrics for 3D Semantic Occupancy Prediction (NeurIPS 2025 · similarity 0.90)
- GaussTR: Foundation Model-Aligned Gaussian Transformer for Self-Supervised 3D Spatial Understanding (CVPR 2025 · similarity 0.90)
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.