Research map / 3D vision / 3D generation
OctreeOcc: Efficient and Multi-Granularity Occupancy Prediction Using Octree Queries
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 OctreeOcc, a camera-based 3D occupancy predictor that builds sparse multi-granularity octree queries initialized from image segmentation priors and iteratively rectified by a learned split predictor, improving mIoU on Occ3D-nuScenes and SemanticKITTI while cutting memory and latency.
Problem
Dense voxel grids for camera-based 3D occupancy prediction are computationally expensive and waste resolution on large objects while losing detail on small ones.
Approach
Replace dense voxel queries with octree queries whose split structure is predicted from image semantic segmentation priors (projected onto voxels) and refined iteratively by an MLP over low-confidence regions, so each region gets a granularity suited to its size.
Key results
- On Occ3D-nuScenes validation, OctreeOcc reaches 44.02 mIoU with camera mask training, versus 43.41 for FB-OCC under the same settings.
- On SemanticKITTI validation, OctreeOcc achieves 44.71 IoU, the best reported IoU, with 13.12 mIoU.
- Octree queries reduce memory and latency relative to dense-query baselines, with the paper claiming a 15%-24% computational overhead reduction.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use octree queries instead of dense queries (ablation) | OctreeOcc · Occ3D-nuScenes validation, reduced input size 0.3x | mIoU: 34.17 -> 34.91 | Helped |
| Add semantic-guided octree initialization (ablation) | OctreeOcc with octree queries · Occ3D-nuScenes validation, reduced input size 0.3x | mIoU: 34.91 -> 36.63 | Helped |
| Add iterative structure rectification (ablation) | OctreeOcc with octree queries · Occ3D-nuScenes validation, reduced input size 0.3x | mIoU: 34.91 -> 35.88 | Helped |
| Use unbalanced assignment in octree initialization (ablation) | OctreeOcc · Occ3D-nuScenes validation, octree structure levels 1 to 2 | mIoU: 45.79 -> 57.34 | Helped |
| Use unbalanced assignment in octree initialization (ablation) | OctreeOcc · Occ3D-nuScenes validation, octree structure levels 2 to 3 | mIoU: 33.60 -> 51.28 | Helped |
| Apply a second rectification iteration instead of stopping after the first (ablation) | OctreeOcc · Occ3D-nuScenes validation, octree structure levels 1 to 2 | mIoU: 60.13 -> 62.27 | Helped |
| Apply a second rectification iteration instead of stopping after the first (ablation) | OctreeOcc · Occ3D-nuScenes validation, octree structure levels 2 to 3 | mIoU: 53.95 -> 56.79 | Helped |
| Replace random query initialization with semantic-guided octree initialization (ablation) | OctreeOcc · Occ3D-nuScenes validation, reduced input size 0.3x | mIoU: 34.91 -> 37.40 | Helped |
| Replace FLoSP voxel features with semantic-guided octree initialization (ablation) | OctreeOcc · Occ3D-nuScenes validation, reduced input size 0.3x | mIoU: 35.72 -> 37.40 | Helped |
| Increase octree depth from 2 to 3 (ablation) | OctreeOcc · Occ3D-nuScenes validation, reduced input size 0.3x | mIoU: 32.02 -> 33.76 | Helped |
| Increase octree depth from 3 to 4 (ablation) | OctreeOcc · Occ3D-nuScenes validation, reduced input size 0.3x | mIoU: 33.76 -> 34.88 | Helped |
| Change selection ratios from 20%, 60% to 10%, 60% (ablation) | OctreeOcc · Occ3D-nuScenes validation, reduced input size 0.3x | mIoU: 37.40 -> 34.47 | Hurt |
Surprising findings
- Octree structure prediction from a randomly initialized query with an MLP lowers mIoU by 1.7 points, showing the semantic prior is essential.
- Applying an OGN-style octree decoder only, without octree encoding, does not reduce computational cost and yields lower mIoU.
Limitations
- Octree depth beyond 3 causes accumulated split-prediction errors and performance degradation.
- Query selection ratios are manually set rather than learned, and binarization of the learned mask needs top-k to avoid pattern collapse.
Experimental setup
Datasets: Occ3D-nuScenes SemanticKITTI
Benchmarks: Occ3D-nuScenes SemanticKITTI
Metrics: mIoU IoU latency memory
Baselines: MonoScene BEVDet BEVFormer BEVStereo TPVFormer OccFormer CTF-Occ RenderOcc BEVDet4D PanoOcc FB-OCC VoxFormer
Builds on: BEVFormer MonoScene FB-OCC OGN
Closest papers by meaning
- OctOcc: High-Resolution 3D Occupancy Prediction with Octree (AAAI 2024 · similarity 0.94)
- LowRankOcc: Tensor Decomposition and Low-Rank Recovery for Vision-based 3D Semantic Occupancy Prediction (CVPR 2024 · similarity 0.89)
- LOMA: Language-assisted Semantic Occupancy Network via Triplane Mamba (AAAI 2025 · similarity 0.89)
- 3D Occupancy Prediction with Low-Resolution Queries via Prototype-aware View Transformation (CVPR 2025 · similarity 0.89)
- AdaOcc: Adaptive 3D Occupancy Prediction for Embodied Tasks (NeurIPS 2026 · similarity 0.89)
- ProtoOcc: Accurate, Efficient 3D Occupancy Prediction Using Dual Branch Encoder-Prototype Query Decoder (AAAI 2025 · similarity 0.88)
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.