Self-positioning Point-based Transformer for Point Cloud Understanding

Official pytorch implementation of "Self-positioning Point-based Transformer for Point Cloud Understanding" (CVPR 2023).

Jinyoung Park ^1* , Sanghyeok Lee ^1* , Sihyeon Kim ¹ , Yunyang Xiong ² , Hyunwoo J. Kim ^1†
1 Korea University, ² Meta Reality Labs

Setup

• Clone repository

git clone https://github.com/mlvlab/SPoTr.git
cd SPoTr

• Install packages with a setup file

bash install.sh

• Dataset

mkdir -p data/S3DIS/
cd data/S3DIS
gdown https://drive.google.com/uc?id=1MX3ZCnwqyRztG1vFRiHkKTz68ZJeHS4Y
tar -xvf s3disfull.tar
cd ../../

cd data && mkdir ShapeNetPart && cd ShapeNetPart
gdown https://drive.google.com/uc?id=1W3SEE-dY1sxvlECcOwWSDYemwHEUbJIS
tar -xvf shapenetcore_partanno_segmentation_benchmark_v0_normal.tar
cd ../../

Run Experiments

If you want pretrained models, you can download pretrained models via this URL

SNPart

• Train

CUDA_VISIBLE_DEVICES='0' python examples/shapenetpart/main.py --cfg cfgs/shapenetpart/spotr.yaml

• Inference

CUDA_VISIBLE_DEVICES='0' python examples/segmentation/main.py --cfg cfgs/shapenetpart/spotr.yaml mode=test --pretrained_path ckpt/ShapeNetPart/ckpt_best.pth

S3DIS

• Train

CUDA_VISIBLE_DEVICES='0' python examples/segmentation/main.py  --cfg cfgs/s3dis/spotr.yaml

• Inference

CUDA_VISIBLE_DEVICES='0' python examples/segmentation/main.py  --cfg cfgs/s3dis/spotr.yaml  mode=test --pretrained_path ckpt/S3DIS/ckpt_best.pth

Acknowledgement

This repo is built upon OpenPoints.

https://github.com/guochengqian/openpoints