This is the implementation of 3D Cascade RCNN: High Quality Object Detection in Point Clouds.
We designed a 3D object detection model on point clouds by:
- Presenting a simple yet effective 3D cascade architecture
- Analyzing the sparsity of the point clouds and using point completeness score to re-weighting training samples. Following is detection results on Waymo Open Dataset.
| Easy Car | Moderate Car | Hard Car | |
|---|---|---|---|
| AP 11 | 90.05 | 86.02 | 79.27 |
| AP 40 | 93.20 | 86.19 | 83.48 |
| Overall Vehicle | 0-30m Vehicle | 30-50m Vehicle | 50m-Inf Vehicle | |
|---|---|---|---|---|
| LEVEL_1 mAP | 76.27 | 92.66 | 74.99 | 54.49 |
| LEVEL_2 mAP | 67.12 | 91.95 | 68.96 | 41.82 |
- Requirements. The code is tested on the following environment:
- Ubuntu 16.04 with 4 V100 GPUs
- Python 3.7
- Pytorch 1.7
- CUDA 10.1
- spconv 1.2.1
- Build extensions
python setup.py develop
Please download the official KITTI dataset and generate data infos by following command:
python -m pcdet.datasets.kitti.kitti_dataset create_kitti_infos tools/cfgs/kitti_dataset.yaml
The folder should be like:
data
├── kitti
│ │── ImageSets
│ │── training
│ │ ├──calib & velodyne & label_2 & image_2
│ │── testing
│ │ ├──calib & velodyne & image_2
| |── kitti_dbinfos_train.pkl
| |── kitti_infos_train.pkl
| |── kitti_infos_val.pkl
The configuration file is in tools/cfgs/3d_cascade_rcnn.yaml, and the training scripts is in tools/scripts.
cd tools
sh scripts/3d-cascade-rcnn.sh
The pre-trained KITTI model is at: model. Run with:
cd tools
sh scripts/3d-cascade-rcnn_test.sh
The evaluation results should be like:
2021-08-10 14:06:14,608 INFO Car AP@0.70, 0.70, 0.70:
bbox AP:97.9644, 90.1199, 89.7076
bev AP:90.6405, 89.0829, 88.4391
3d AP:90.0468, 86.0168, 79.2661
aos AP:97.91, 90.00, 89.48
Car AP_R40@0.70, 0.70, 0.70:
bbox AP:99.1663, 95.8055, 93.3149
bev AP:96.3107, 92.4128, 89.9473
3d AP:93.1961, 86.1857, 83.4783
aos AP:99.13, 95.65, 93.03
Car AP@0.70, 0.50, 0.50:
bbox AP:97.9644, 90.1199, 89.7076
bev AP:98.0539, 97.1877, 89.7716
3d AP:97.9921, 90.1001, 89.7393
aos AP:97.91, 90.00, 89.48
Car AP_R40@0.70, 0.50, 0.50:
bbox AP:99.1663, 95.8055, 93.3149
bev AP:99.1943, 97.8180, 95.5420
3d AP:99.1717, 95.8046, 95.4500
aos AP:99.13, 95.65, 93.03
@article{cascade3d,
title={3D Cascade RCNN: High Quality Object Detection in Point Clouds},
author={Cai, Qi and Pan, Yingwei and Yao, Ting and Mei, Tao},
journal={IEEE Transactions on Image Processing},
year={2022},
publisher={IEEE}
}
The code is built on OpenPCDet and Voxel R-CNN.