Research map / Architectures and efficiency / Attention and transformers
OccMamba: Semantic Occupancy Prediction with State Space Models
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 OccMamba, the first Mamba-based network for outdoor semantic occupancy prediction, using a height-prioritized 2D Hilbert reordering and hierarchical Mamba plus local context processor on top of M-CONet.
Problem
Semantic occupancy prediction from camera and LiDAR involves huge dense voxel grids, and transformer-based models have quadratic cost; converting 3D voxels to 1D sequences for Mamba breaks spatial adjacency.
Approach
Replace transformer/CNN occupancy encoders with a hierarchical Mamba module that processes full voxel grids, using a height-prioritized 2D Hilbert expansion to reorder 3D voxels into 1D sequences while preserving spatial proximity, plus a windowed local context processor.
Key results
- On OpenOccupancy validation, OccMamba-384 outperforms Co-Occ by 5.1% IoU and 4.3% mIoU.
- On SemanticKITTI test, OccMamba-128 reaches 24.6 mIoU, 0.2 above Co-Occ (24.4).
- Height-prioritized 2D Hilbert expansion gives 25.2 mIoU versus 24.5 for XYZ sequence ordering and 24.8 for 3D Hilbert; OccMamba-128 inference takes about 35% of M-CONet's time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the CNN-based encoder with OccMamba | semantic occupancy prediction · OpenOccupancy validation v0.0 | IoU: 30.6 -> 35.7 | Helped |
| replace the CNN-based encoder with OccMamba | semantic occupancy prediction · OpenOccupancy validation v0.0 | mIoU: 21.9 -> 26.2 | Helped |
| use OccMamba-128 instead of Co-Occ | semantic occupancy prediction · SemanticKITTI test | mIoU: 24.4 -> 24.6 | Helped |
| use OccMamba-128 instead of JS3C-Net | semantic occupancy prediction · SemanticPOSS validation | mIoU: 22.7 -> 23.4 | Helped |
| use height-prioritized 2D Hilbert expansion instead of XYZ sequence reordering (ablation) | OccMamba-128 · OpenOccupancy validation v0.0 | mIoU: 24.5 -> 25.2 | Helped |
| use height-prioritized 2D Hilbert expansion instead of ZXY sequence reordering (ablation) | OccMamba-128 · OpenOccupancy validation v0.0 | mIoU: 24.7 -> 25.2 | Helped |
| use height-prioritized 2D Hilbert expansion instead of 3D Hilbert reordering (ablation) | OccMamba-128 · OpenOccupancy validation v0.0 | mIoU: 24.8 -> 25.2 | Helped |
| add a hierarchical Mamba module instead of using the CNN occupancy encoder (ablation) | M-CONet · OpenOccupancy validation v0.0 | mIoU: 20.1 -> 25.0 | Helped |
| add a local context processor (ablation) | OccMamba-128 with hierarchical Mamba module · OpenOccupancy validation v0.0 | mIoU: 25.0 -> 25.2 | Helped |
| increase the Mamba feature dimension from 128 to 384 (ablation) | OccMamba · OpenOccupancy validation v0.0 | mIoU: 25.2 -> 26.2 | Helped |
| replace the 3D voxel decoder with a hierarchical Mamba module (ablation) | MonoScene · SemanticKITTI validation | mIoU: 11.1 -> 11.7 | Helped |
| add a local context processor (ablation) | MonoScene with hierarchical Mamba module · SemanticKITTI validation | mIoU: 11.7 -> 11.9 | Helped |
Surprising findings
- Reordering schemes that prioritize the z-axis outperform the standard XYZ and 3D Hilbert orderings, attributed to the flattened cuboid shape of scenes.
Limitations
- Mamba-based design is evaluated mainly against CNN/transformer occupancy models built on M-CONet; the gains over Co-Occ on SemanticKITTI are small (0.2 mIoU).
- SemanticKITTI and SemanticPOSS comparisons exclude methods using tricks such as label rectification, distillation, or multi-frame input.
Experimental setup
Datasets: OpenOccupancy SemanticKITTI SemanticPOSS
Benchmarks: OpenOccupancy SemanticKITTI SemanticPOSS
Metrics: IoU mIoU GPU memory usage inference time
Baselines: M-CONet Co-Occ MonoScene TPVFormer SparseOcc JS3C-Net LMSCNet AICNet
Builds on: Mamba M-CONet Hilbert curve Point Mamba
Closest papers by meaning
- LOMA: Language-assisted Semantic Occupancy Network via Triplane Mamba (AAAI 2025 · similarity 0.92)
- LowRankOcc: Tensor Decomposition and Low-Rank Recovery for Vision-based 3D Semantic Occupancy Prediction (CVPR 2024 · similarity 0.90)
- OctOcc: High-Resolution 3D Occupancy Prediction with Octree (AAAI 2024 · similarity 0.90)
- Achieving Speed-Accuracy Balance in Vision-based 3D Occupancy Prediction via Geometric-Semantic Disentanglement (AAAI 2025 · similarity 0.89)
- OPUS: Occupancy Prediction Using a Sparse Set (NeurIPS 2024 · similarity 0.89)
- CSV-Occ: Fusing Multi-frame Alignment for Occupancy Prediction with Temporal Cross State Space Model and Central Voting Mechanism (ICML 2025 · similarity 0.89)
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.