This repo contains the implementations of Object DGCNN (https://arxiv.org/abs/2110.06923) and DETR3D (https://arxiv.org/abs/2110.06922). Our implementations are built on top of MMdetection3D.
- Follow the mmdet3d to process the data.
-
Downloads the pretrained backbone weights to pretrained/
-
For example, to train Object-DGCNN with pillar on 8 GPUs, please use
tools/dist_train.sh projects/configs/obj_dgcnn/pillar.py 8
- Download the weights accordingly.
| Backbone | mAP | NDS | Download |
|---|---|---|---|
| DETR3D, ResNet101 w/ DCN | 34.7 | 42.2 | model | log |
| above, + CBGS | 34.9 | 43.4 | model | log |
| DETR3D, VoVNet on trainval, evaluation on test set | 41.2 | 47.9 | model | log |
| Backbone | mAP | NDS | Download |
|---|---|---|---|
| Object DGCNN, pillar | 53.2 | 62.8 | model | log |
| Object DGCNN, voxel | 58.6 | 66.0 | model | log |
- To test, use
tools/dist_test.sh projects/configs/obj_dgcnn/pillar_cosine.py /path/to/ckpt 8 --eval=bbox
If you find this repo useful for your research, please consider citing the papers
@inproceedings{ obj-dgcnn, title={Object DGCNN: 3D Object Detection using Dynamic Graphs}, author={Wang, Yue and Solomon, Justin M.}, booktitle={2021 Conference on Neural Information Processing Systems ({NeurIPS})}, year={2021} } @inproceedings{ detr3d, title={DETR3D: 3D Object Detection from Multi-view Images via 3D-to-2D Queries}, author={Wang, Yue and Guizilini, Vitor and Zhang, Tianyuan and Wang, Yilun and Zhao, Hang and and Solomon, Justin M.}, booktitle={The Conference on Robot Learning ({CoRL})}, year={2021} } 