fudan-zvg / DeepInteraction

[NeurIPS 2022] DeepInteraction: 3D Object Detection via Modality Interaction

DeepInteraction: 3D Object Detection via Modality Interaction

Paper

DeepInteraction: 3D Object Detection via Modality Interaction,
Zeyu Yang, Jiaqi Chen, Zhenwei Miao, Wei Li, Xiatian Zhu, Li Zhang
NeurIPS 2022

News

• (2022/6/27) DeepInteraction-e ranks first on nuScenes among all solutions.
• (2022/6/26) DeepInteraction-large ranks first on nuScenes among all non-ensemble solutions.
• (2022/5/18) DeepInteraction-base ranks first on nuScenes among all solutions that do not use test-time augmentation and model ensemble.

Results

3D Object Detection (on nuScenes test)

Model	Modality	mAP	NDS
DeepInteraction-e	C+L	75.74	76.34
DeepInteraction-large	C+L	74.12	75.52
DeepInteraction-base	C+L	70.78	73.43

3D Object Detection (on nuScenes val)

Model	Modality	mAP	NDS	Checkpoint
DeepInteraction-base	C+L	69.85	72.63	Fusion_0075_refactor.pth

Get Started

Environment

This implementation is build upon mmdetection3d, please follow the steps in install.md to prepare the environment.

Data

Please follow the official instructions of mmdetection3d to process the nuScenes dataset.(https://mmdetection3d.readthedocs.io/en/latest/datasets/nuscenes_det.html)

Pretrained

Downloads the pretrained backbone weights to pretrained/

Train & Test

# train with 8 GPUs
tools/dist_train.sh projects/configs/nuscenes/Fusion_0075_refactor.py 8
# test with 8 GPUs
tools/dist_test.sh projects/configs/nuscenes/Fusion_0075_refactor.py ${CHECKPOINT_FILE} 8 --eval=bbox

Acknowledgement

Many thanks to the following open-source projects:

• mmdetection3d
• detectron2
• transfusion

Reference

@inproceedings{yang2022deepinteraction,
  title={DeepInteraction: 3D Object Detection via Modality Interaction},
  author={Yang, Zeyu and Chen, Jiaqi and Miao, Zhenwei and Li, Wei and Zhu, Xiatian and Zhang, Li},
  booktitle={NeurIPS},
  year={2022}
}