This is the official PyTorch implementation of the papers :
Learning Inter-Superpoint Affinity for Weakly Supervised 3D Instance Segmentation (ACCV 2022) [arxiv]
by Linghua Tang, Le Hui, and Jin Xie.
1) Requirements
2) Anaconda Virtual Environment
conda create -n 3DWSIS python=3.7
conda activate 3DWSIS
3) Clone the repository.
git clone https://github.com/fpthink/3D-WSIS.git --recursive
4) Install the requirements.
cd 3DWSIS
pip install -r requirements.txt
conda install -c bioconda google-sparsehash
5) Install spconv
and pointgroup_ops
Please refer to PointGroup to install.
Please refer to the ScanNetV2.md
in data/ScanNetV2
to process data.
Please set $ScanNetV2_DATA
on Line 29
of config/ScanNet_v2_3D_WSIS.yaml
.
CUDA_VISIBLE_DEVICES=0 python train_scannetv2.py --config config/ScanNet_v2_3D_WSIS.yaml
CUDA_VISIBLE_DEVICES=0 python test_scannetv2.py --config config/ScanNet_v2_3D_WSIS.yaml --pretrain log/ScanNet_v2_3D_WSIS/epoch_00120_whole_scene.pth
Please refer to the S3DIS.md
in data/S3DIS
to process data.
Please set $S3DIS_DATA/data
on Line 29
of config/S3DIS_Area5_3D_WSIS.yaml
.
CUDA_VISIBLE_DEVICES=0 python train_s3dis.py --config config/S3DIS_Area5_3D_WSIS.yaml
CUDA_VISIBLE_DEVICES=0 python test_s3dis.py --config config/S3DIS_Area5_3D_WSIS.yaml --pretrain log/S3DIS_Area5_3D_WSIS/epoch_00300_whole_scene.pth
Its performance on ScanNet-v2 validation set is 29.8/48.4/67.7 in terms of mAP/mAP50/mAP25.
Its performance on S3DIS Area5 set is 22.4/35.2/47.2/43.2/44.7/51.8/41.3 in terms of mAP/mAP50/mAP25/mCov/mWCov/mPrec/mRec.
Note : Due to the randomness of weak label generation, the results of network training fluctuate slightly.
This repo is built upon several repos, e.g., PointGroup, SSTNet, HAIS, spconv and ScanNet.
If you find this work useful in your research, please cite:
@inproceedings{tang20223dwsis,
author = {Tang, Linghua and Hui, Le and Xie, Jin},
title = {Learning Inter-Superpoint Affinity for Weakly Supervised 3D Instance Segmentation},
booktitle = {ACCV},
year = {2022},
}
@inproceedings{hui2022graphcut,
author = {Hui, Le and Tang, Linghua and Shen, Yaqi and Xie, Jin and Yang, Jian},
title = {Learning Superpoint Graph Cut for 3D Instance Segmentation},
booktitle = {NeurIPS},
year = {2022},
}