Research map / Point clouds / Segmentation and detection
UniMamba: Unified Spatial-Channel Representation Learning with Group-Efficient Mamba for LiDAR-based 3D Object Detection
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 UniMamba, a LiDAR 3D backbone combining 3D submanifold convolution, complementary Z-order serialization, and a channel-grouped local-global Mamba aggregator for 3D object detection.
Problem
Transformer-based LiDAR detectors serialize voxels into 1D sequences, losing spatial structure and limiting receptive fields due to window grouping; naive Mamba adaptation loses spatial locality and lacks spatial diversity.
Approach
Integrates submanifold 3D convolution for locality before serialization, uses bidirectional Z-order (X- and Y-primary) serialization, and splits channels into groups so that some run a global SSM over all voxels and others run a local windowed SSM, aggregated in a multi-head fashion.
Key results
- Achieves 70.2 mAP and 74.0 NDS on nuScenes test set, the first LiDAR method above 70 mAP there.
- Reaches 76.13 L2 mAP on Waymo validation, surpassing HEDNet (75.30) and LION-Mamba-L (75.90).
- Obtains 42.0 mAP on Argoverse 2 validation, 0.5 above LION-Mamba (41.5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace SAFDNet with UniMamba | LiDAR 3D detection · nuScenes test | mAP: 68.3 -> 70.2 | Helped |
| Replace LION-Mamba-L with UniMamba | LiDAR 3D detection · Waymo Open validation, ALL (L2) | mAP: 75.90 -> 76.13 | Helped |
| Replace LION-Mamba with UniMamba | LiDAR 3D detection · Argoverse 2 val | mAP: 41.5 -> 42.0 | Helped |
| Add spatial locality modeling (SLM) (ablation) | UniMamba with complementary Z-order serialization · nuScenes val | mAP: 67.6 -> 67.9 | Helped |
| Replace single-direction Z-order with complementary Z-order (ablation) | UniMamba with SLM · nuScenes val | mAP: +0.3 mAP | Helped |
| Use Hilbert instead of Z-order serialization (ablation) | UniMamba without SLM · nuScenes val | mAP: 67.2 -> 67.4 | Helped |
| Add SLM to random serialization (ablation) | UniMamba · nuScenes val | mAP: 66.2 -> 67.4 | Helped |
| Add both LSE and GSE rather than the baseline without either (ablation) | Transfusion-L backbone · nuScenes | mAP: 65.7 -> 68.5 | Helped |
| Use both local and global encoders rather than LSE only or GSE only (ablation) | Transfusion-L backbone · nuScenes | mAP: 67.7 -> 68.5; 67.8 -> 68.5 | Helped |
| Use 4 balanced channel groups instead of 8 balanced groups (ablation) | UniMamba LGSA · nuScenes | mAP: 68.2 -> 68.5 | Helped |
| Use balanced LSE/GSE groups instead of 4 groups with a 1:3 or 3:1 split (ablation) | UniMamba LGSA · nuScenes | mAP: 68.4 -> 68.5; 68.3 -> 68.5 | Helped |
| Aggregate LSE and GSE with channel grouping instead of sequentially (ablation) | UniMamba LGSA · nuScenes | mAP: 68.0 -> 68.5 | Helped |
Surprising findings
- Random serialization with spatial locality modeling reaches 67.4 mAP, nearly matching Hilbert and Z-order curves, suggesting the convolution provides most of the locality benefit.
Limitations
- Gains over strong SpCNN and Mamba baselines are modest (often under 1 mAP) on several benchmarks.
- Built on top of LION and SAFDNet pipelines, so the backbone's standalone contribution depends on those detector configurations.
Experimental setup
Datasets: nuScenes Waymo Open Dataset Argoverse 2
Benchmarks: nuScenes Waymo Open Dataset Argoverse 2
Metrics: mAP NDS L1/L2 mAP mAPH latency FLOPs parameters
Baselines: CenterPoint Transfusion-L VoxelNeXt SAFDNet LION-Mamba DSVT HEDNet VoxelMamba FlatFormer FSDv2
Builds on: LION Vim FlatFormer HEDNet SAFDNet Mamba
Closest papers by meaning
- Voxel Mamba: Group-Free State Space Models for Point Cloud based 3D Object Detection (NeurIPS 2024 · similarity 0.89)
- 3DET-Mamba: Causal Sequence Modelling for End-to-End 3D Object Detection (NeurIPS 2024 · similarity 0.88)
- GBlobs: Explicit Local Structure via Gaussian Blobs for Improved Cross-Domain LiDAR-based 3D Object Detection (CVPR 2025 · similarity 0.87)
- HOTFormerLoc: Hierarchical Octree Transformer for Versatile Lidar Place Recognition Across Ground and Aerial Views (CVPR 2025 · similarity 0.86)
- Spatial-Mamba: Effective Visual State Space Models via Structure-Aware State Fusion (ICLR 2025 · similarity 0.86)
- PointMamba: A Simple State Space Model for Point Cloud Analysis (NeurIPS 2024 · similarity 0.86)
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.