Suppress-and-Refine Framework for End-to-End 3D Object Detection

This repo is the official implementation of "Suppress-and-Refine Framework for End-to-End 3D Object Detection".

A simple, fast, efficient and end-to-end 3D object detector without NMS.

Getting Started

Main results

ScanNet V2

Method	backbone	mAP@0.25	mAP@0.5	Runtime (FPS)	Ckpt
VoteNet	PointNet++	62.9	39.9	10.8	-
H3DNet	4xPointNet++	67.2	48.1	4.4	-
MLCVNet	PointNet++	64.5	41.4	6.7	-
BRNet	PointNet++	66.1	50.9	8.7	-
Group-Free	PointNet++	67.3	48.9	7.1	-
Ours	PointNet++	66.2	53.5	13.5	model_ckpt

SUN RGB-D

Method	backbone	mAP@0.25	mAP@0.5	Ckpt
VoteNet	PointNet++	59.1	35.8	-
H3DNet	4xPointNet++	60.1	39.0	-
MLCVNet	PointNet++	59.8	-	-
BRNet	PointNet++	61.1	43.7	-
Group-Free	PointNet++	63.0	45.2	-
Ours	PointNet++	60.0	44.7	model_ckpt

The FPS is tested on a V100 GPU.

Quick start

Installation

This repository is based on mmdetection3d, please follow this page for installation guidance.

Reproduce our results on SCANNET and SUNRGBD

For SCANNET.

CUDA_VISIBLE_DEVICES=0,1 PORT=29600 ./tools/dist_train.sh configs/sr/scannet_baseline.py 2

For SUNRGBD

CUDA_VISIBLE_DEVICES=0,1 PORT=29600 ./tools/dist_train.sh configs/sr/sunrgbd_baseline.py 2

Evaluation

Please first download the ckpt from the ckpt link provided above.

Then for SCANNET.

./tools/dist_test.sh configs/sr/scannet_baseline.py epoch_30.pth 2 --eval mAP

For SUNRGBD

./tools/dist_test.sh configs/sr/sunrgbd_baseline.py epoch_33.pth 4 --eval mAP

Acknowledgement

Our code is based on wonderful mmdetection3d. Very apperciate their works!

Citation

If you find this project useful in your research, please consider cite:

@article{liu2021suppress,
  title={Suppress-and-Refine Framework for End-to-End 3D Object Detection},
  author={Liu, Zili and Xu, Guodong and Yang, Honghui and Chen, Minghao and Wu, Kuoliang and Yang, Zheng and Liu, Haifeng and Cai, Deng},
  journal={arXiv preprint arXiv:2103.10042},
  year={2021}
}