We incorporate PIoU loss in projects/MonoRCNN/lib/att_head.py.
To use only L1 loss, delete lines 321-326 in att_head.py and fix loss_kpt_iou to be 0.
config files for L1 and L1+PIoU loss are in projects/MonoRCNN/config/.
MonoRCNN is a monocular 3D object detection method for autonomous driving, published at ICCV 2021 and WACV 2023. This project is an implementation of MonoRCNN.
- Video demo (KITTI)
- Paper (ICCV 2021)
- Paper (WACV 2023)
- KITTI benchmark (ICCV 2021)
- KITTI benchmark (WACV 2023)
- KITTI
- WAYMO
- Python 3.6
- PyTorch 1.5.0
- Detectron2 0.1.3
Please use the Detectron2 included in this project. To ignore fully occluded objects during training, build.py, rpn.py, and roi_heads.py have been modified.
Organize the downloaded files as follows:
├── projects
│ ├── MonoRCNN
│ │ ├── output
│ │ │ ├── model
│ │ │ ├── log.txt
│ │ │ ├── ...
cd projects/MonoRCNN
./main.py --config-file config/MonoRCNN_KITTI.yaml --num-gpus 1 --resume --eval-only
Set VISUALIZE as True to visualize 3D object detection results (saved in output/evaluation/test/visualization).
cd projects/MonoRCNN
./main.py --config-file config/MonoRCNN_KITTI.yaml --num-gpus 1
If you find this project useful in your research, please cite:
@inproceedings{MonoRCNN_ICCV21,
title = {Geometry-based Distance Decomposition for Monocular 3D Object Detection},
author = {Xuepeng Shi and Qi Ye and
Xiaozhi Chen and Chuangrong Chen and
Zhixiang Chen and Tae-Kyun Kim},
booktitle = {ICCV},
year = {2021},
}
@inproceedings{MonoRCNN_WACV23,
title = {Multivariate Probabilistic Monocular 3D Object Detection},
author = {Xuepeng Shi and Zhixiang Chen and Tae-Kyun Kim},
booktitle = {WACV},
year = {2023},
}