A general 3D Object Detection codebase in PyTorch.
Det3D is the first 3D Object Detection toolbox which provides off the box implementations of many 3D object detection algorithms such as PointPillars, SECOND, PIXOR, etc, as well as state-of-the-art methods on major benchmarks like KITTI(ViP) and nuScenes(CBGS). Key features of Det3D include the following aspects:
- Multi Datasets Support: KITTI, nuScenes, Lyft
- Point-based and Voxel-based model zoo
- State-of-the-art performance
- DDP & SyncBN
Please refer to INSTALATION.md.
Please refer to GETTING_STARTED.md.
| mAP | mATE | mASE | mAOE | mAVE | mAAE | NDS | ckpt | |
|---|---|---|---|---|---|---|---|---|
| CBGS | 49.9 | 0.335 | 0.256 | 0.323 | 0.251 | 0.197 | 61.3 | link |
| PointPillar | 41.8 | 0.363 | 0.264 | 0.377 | 0.288 | 0.198 | 56.0 | link |
The original model and prediction files are available in the CBGS README.
Second on KITTI(val) Dataset
car AP @0.70, 0.70, 0.70:
bbox AP:90.54, 89.35, 88.43
bev AP:89.89, 87.75, 86.81
3d AP:87.96, 78.28, 76.99
aos AP:90.34, 88.81, 87.66
PointPillars on KITTI(val) Dataset
car [email protected], 0.70, 0.70:
bbox AP:90.63, 88.86, 87.35
bev AP:89.75, 86.15, 83.00
3d AP:85.75, 75.68, 68.93
aos AP:90.48, 88.36, 86.58
- Models
- VoxelNet
- SECOND
- PointPillars
- Features
- Multi task learning & Multi-task Learning
- Distributed Training and Validation
- SyncBN
- Flexible anchor dimensions
- TensorboardX
- Checkpointer & Breakpoint continue
- Self-contained visualization
- Finetune
- Multiscale Training & Validation
- Rotated RoI Align
-
To Be Released
- CGBS on Lyft(val) Dataset
-
Models
- PointRCNN
- PIXOR
- Support Waymo Dataset.
- Add other 3D detection / segmentation models, such as VoteNet, STD, etc.
Det3D is released under the Apache licenes.
Det3D is a derivative codebase of CBGS, if you find this work useful in your research, please consider cite:
@article{zhu2019class,
title={Class-balanced Grouping and Sampling for Point Cloud 3D Object Detection},
author={Zhu, Benjin and Jiang, Zhengkai and Zhou, Xiangxin and Li, Zeming and Yu, Gang},
journal={arXiv preprint arXiv:1908.09492},
year={2019}
}
