ASSANet

arXiv | Supplementary | New version of ASSANet Implementation

This is official repository of NeurIPS 2021 spotlight paper: ASSANet: An Anisotropical Separable Set Abstraction forEfficient Point Cloud Representation Learning.

This repo is deprecated Please refer to PointNeXt for the latest code of ASSANet

News

• 2022/06/10: This repo is not maintained anymore. Although one can still use the repo for reproducing our results in paper, we highly suggest one to use our re-implemented version hosted in PointNeXt&OpenPoints Lib, where one can reproduce ASSANet with eased coding.

Installation

Datasets

Scene Segmentation on S3DIS Download the S3DIS dataset from here (4.8 GB). You only need to download the file named Stanford3dDataset_v1.2.zip, unzip and move (or link) it to data/S3DIS/Stanford3dDataset_v1.2.

The file structure should look like:

<pt-code-root>
├── cfgs
│   └── s3dis
├── data
│   └── S3DIS
│       └── Stanford3dDataset_v1.2
│           ├── Area_1
│           ├── Area_2
│           ├── Area_3
│           ├── Area_4
│           ├── Area_5
│           └── Area_6
├── init.sh
├── datasets
├── function
├── models
├── ops
└── utils

Environment install

Make sure install CUDA 11.1

source init.sh

Usage

Training

• ASSA-Net:

  python -m torch.distributed.run --nnodes 1 --nproc_per_node 1 function/main_s3dis_dist.py --cfg cfgs/s3dis/assanet.yaml

• SSA-Net (L):

  python -m torch.distributed.run --nnodes 1 --nproc_per_node 1 function/main_s3dis_dist.py --cfg cfgs/s3dis/assanet_scale.yaml model.width 128 model.depth 3

  Note: we trained all models using the default hyperparameters and using only one GPU (32G V100). ASSANet is also trainable in one GTX2080Ti. ASSANet (L) can be trained using multiple GTX2080Ti.

Evaluating

For evaluation, 1 GPU is recommended.

Evaluate ASSA-Net:

python -m torch.distributed.run --nnodes 1 --nproc_per_node 1 \
    function/main_s3dis_dist.py mode=test wandb.use_wandb=False --cfg cfgs/s3dis/assanet.yaml --load_path /path/to/the/ckpt 

Evaluate ASSA-Net(L):

python -m torch.distributed.run --nnodes 1 --nproc_per_node 1 \
    function/main_s3dis_dist.py mode=test wandb.use_wandb=False --cfg cfgs/s3dis/assanet_scale.yaml model.width=128 model.depth=3 --load_path /path/to/the/ckpt

Model Zoo

Model (S3DIS)	Paper (mIoU, with voting)	Reproduce (with voting / without voting)	Ckpt & Logs
ASSANet	63.0	62.9 / 61.8	Google Drive
ASSANet-L	66.8	66.7 / 64. 3	Google Drive

Acknowledge

This code is built upon Closer Look at 3D.

Cite

@inproceedings{qian2021assanet,
  title={ASSANet: An Anisotropical Separable Set Abstraction for Efficient Point Cloud Representation Learning},
  author={Qian, Guocheng and Hammoud, Hasan and Li, Guohao and 
          Thabet, Ali and Ghanem, Bernard},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2021}
}