Research map / 3D vision / 3D generation
OctOcc: High-Resolution 3D Occupancy Prediction with Octree
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 OctOcc, an octree-based network that runs 3D-2D cross-attention only in potentially occupied voxel regions at multiple resolutions, plus a 2D image feature filter, to enable memory-efficient 200x200x16 semantic occupancy prediction.
Problem
Existing vision-based 3D semantic occupancy prediction methods are limited to low-resolution 3D voxel features because high-resolution dense cross-attention is too memory-expensive.
Approach
Exploit the sparsity of occupied space: a hierarchical octree predicts occupied coarse voxels via binary occupancy supervision, and only their octant children get finer cross-attention queries; a 2D convolutional filter additionally masks irrelevant image regions (sky, far objects) using dilated LiDAR depth masks.
Key results
- OctOcc reaches 31.52 mIoU on nuScenes occupancy validation at 200x200x16, versus 27.03 for SurroundOcc and 28.53 for Occ3D under the same labels.
- On SemanticKITTI, OctOcc achieves 14.59 mIoU and 44.02 IoU, exceeding OccFormer's 13.46 mIoU.
- The octree design raises mIoU by 2.01 over the dense 100x100x8 variant at 23.4G memory, whereas the dense 200x200x16 variant exceeds 32G.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use OctOcc instead of Occ3D | 3D semantic occupancy prediction · nuScenes validation | mIoU: 28.53 -> 31.52 | Helped |
| use OctOcc instead of OccFormer | semantic scene completion · SemanticKITTI validation | mIoU: 13.46 -> 14.59 | Helped |
| use OctOcc instead of VoxFormer | semantic scene completion · SemanticKITTI validation | IoU: 44.15 -> 44.02 | No clear effect |
| add the octree structure to the non-octree variant (ablation) | OctOcc · nuScenes validation | mIoU: +2.01 | Helped |
| add 2D image feature filtering (ablation) | OctOcc · nuScenes validation, 8 epochs | mIoU: 29.8 -> 31.1(+1.3) | Helped |
| add 2D image feature filtering (ablation) | OctOcc · nuScenes validation, 24 epochs | mIoU: 31.23 -> 31.52(+0.3) | Helped |
| add modified Generalized Dice loss alongside cross-entropy (ablation) | OctOcc · nuScenes validation | mIoU: 29.32 -> 31.52 | Helped |
| modify the Generalized Dice loss (ablation) | OctOcc with cross-entropy loss · nuScenes validation | mIoU: 30.64 -> 31.52 | Helped |
Surprising findings
- The original Generalized Dice loss degrades performance (30.64 vs 29.32 CE only baseline gains are reduced relative to the modified version), so the weighting change is necessary.
Limitations
- Octree top-K thresholds (625, 3000, 15000) are hand-set hyperparameters.
- Memory still reaches 23.4G, which is not negligible.
- Occupied-region definition uses a 10% threshold chosen by statistical analysis.
Experimental setup
Datasets: nuScenes Occ3D SemanticKITTI
Benchmarks: nuScenes SemanticKITTI
Metrics: mIoU IoU memory
Baselines: BEVFormer TPVFormer SurroundOcc Occ3D OccFormer MonoScene VoxFormer CE + Sem&Geo loss
Builds on: BEVFormer SurroundOcc Generalized Dice loss MonoScene
Closest papers by meaning
- OctreeOcc: Efficient and Multi-Granularity Occupancy Prediction Using Octree Queries (NeurIPS 2024 · similarity 0.94)
- Achieving Speed-Accuracy Balance in Vision-based 3D Occupancy Prediction via Geometric-Semantic Disentanglement (AAAI 2025 · similarity 0.92)
- LowRankOcc: Tensor Decomposition and Low-Rank Recovery for Vision-based 3D Semantic Occupancy Prediction (CVPR 2024 · similarity 0.92)
- CSV-Occ: Fusing Multi-frame Alignment for Occupancy Prediction with Temporal Cross State Space Model and Central Voting Mechanism (ICML 2025 · similarity 0.91)
- 3D Occupancy Prediction with Low-Resolution Queries via Prototype-aware View Transformation (CVPR 2025 · similarity 0.91)
- ViPOcc: Leveraging Visual Priors from Vision Foundation Models for Single-View 3D Occupancy Prediction (AAAI 2025 · similarity 0.91)
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.