Voxel Mamba: Group-Free State Space Models for Point Cloud based 3D Object Detection
This repo is the official implementation of the paper Voxel Mamba: Group-Free State Space Models for Point Cloud based 3D Object Detection. Our Voxel Mamba achieves state-of-the-art performance on Waymo and nuScene datasets.
-[24-6-18] Voxel Mamba released on arxiv
Validation set
Model | mAPH_L1 | mAPH_L2 | Veh_L1 | Veh_L2 | Ped_L1 | Ped_L2 | Cyc_L1 | Cyc_L2 | Log |
---|---|---|---|---|---|---|---|---|---|
Voxel Mamba | 79.6 | 73.6 | 80.8/80.3 | 72.6/72.2 | 85.0/80.8 | 77.7/73.6 | 78.6/77.6 | 75.7/74.8 | Log |
Test set
Model | mAPH_L1 | mAPH_L2 | Veh_L1 | Veh_L2 | Ped_L1 | Ped_L2 | Cyc_L1 | Cyc_L2 | Leaderboard |
---|---|---|---|---|---|---|---|---|---|
Voxel Mamba | 79.6 | 74.3 | 84.4/84.0 | 77.0/76.6 | 84.8/80.6 | 79.0/74.9 | 75.4/74.3 | 72.6/71.5 | leaderboard |
Validation set
Model | mAP | NDS | mATE | mASE | mAOE | mAVE | mAAE | ckpt | Log |
---|---|---|---|---|---|---|---|---|---|
Voxel Mamba | 67.5 | 71.9 | 26.7 | 25.0 | 25.8 | 21.8 | 18.9 | ckpt | Log |
Test set
Model | mAP | NDS | mATE | mASE | mAOE | mAVE | mAAE | Leaderboard | Submission |
---|---|---|---|---|---|---|---|---|---|
Voxel Mamba | 69.0 | 73.0 | 24.3 | 23.0 | 30.9 | 23.7 | 13.3 | leaderboard | Submission |
Voxel Mamba's result on Waymo compared with other leading methods. All the experiments are evaluated on an NVIDIA A100 GPU with the same environment. We hope that our Voxel Mamba can provide a potential group-free solution for efficiently handling sparse point clouds for 3D tasks.
Please refer to INSTALL.md for installation.
Please follow the instructions from OpenPCDet. We adopt the same data generation process.
cd data
mkdir hilbert
python ./tools/hilbert_curves/create_hilbert_curve_template.py
# multi-gpu training
cd tools
bash scripts/dist_train.sh 8 --cfg_file <CONFIG_FILE>
# multi-gpu testing
cd tools
bash scripts/dist_test.sh 8 --cfg_file <CONFIG_FILE> --ckpt <CHECKPOINT_FILE>
Please consider citing our work as follows if it is helpful.
Voxel Mamba is based on OpenPCDet and DSVT.
We also thank the Centerpoint, TransFusion, OctFormer, and HEDNet authors for their efforts.