AADNet

Point Cloud Semantic Segmentation with Sparse and Inhomogeneous Annotations

Zhiyi Pan, Nan Zhang, Wei Gao, Shan Liu, Ge Li.

Installation Guide

Prerequisites

Before installing our program, please install PointNeXt first.

Installation steps

1. Move all the files of this project to the root directory of PointNeXt.
2. Run the install script install_AAD.sh, or replace the corresponding files according to the guide.

Usage for S3DIS

Dataset

The presampling is just to collect all point clouds, area by area and room by room, following PointNeXt. We provide homogeneous and inhomogeneous annotations with 10%, 1%, 0.1% and 0.01% label rates.

mkdir -p data/S3DIS/
cd data/S3DIS
gdown https://drive.google.com/u/2/uc?id=1uMA58XjKjkmxwIq3dyIrMCIFWnZ0j_41
unzip s3dis.zip

Training

For example, train PointNeXt-l with 1% homogeneous sparse annotations

CUDA_VISIBLE_DEVICES=0 python examples/segmentation/main.py --cfg cfgs/s3dis/pointnext-l-uni-1.yaml

Test on Area 5

CUDA_VISIBLE_DEVICES=0 python examples/segmentation/main.py cfgs/s3dis/<YOUR_CONFIG> wandb.use_wandb=False mode=test --pretrained_path <YOUR_CHECKPOINT_PATH>

Usage for ScanNet

Dataset

For homogeneous annotations, we provide three settings include 1% label rate, 20 points per scene and 20 points per scene (OTOC setting). While for inhomogeneous annotations, we provide annotations at 10% and 1% label rates. You can download our preprocessed ScanNet dataset for weak supervision as follows:

cd data
gdown https://drive.google.com/u/2/uc?id=1OUNKjw84hI3Xj9ucxRdNEpjGjt4Fbrg8
tar -xvf ScanNet.tar

Training

For example, train PointNeXt-l with 1% homogeneous sparse annotations

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python examples/segmentation/main.py --cfg cfgs/scannet/pointnext-l-1.yaml

Val

CUDA_VISIBLE_DEVICES=0  python examples/segmentation/main.py --cfg cfgs/scannet/<YOUR_CONFIG> mode=test dataset.test.split=val --pretrained_path <YOUR_CHECKPOINT_PATH>

Algorithm implementation

The algorithm implementation in the program can be found in the openpoints/dataset/data_util.py and openpoints/loss/bulid.py, respectively:

• Label-aware point cloud downsampling strategy (LaDS) is implemented by rewriting the voxelize function (line 128 ~ line 153) in openpoints/dataset/data_util.py.
• Multiplicative dynamic entropy for asynchronous training (MDE-AT) is implemented by constructing the AsynchronousCrossEntropy function (line 13 ~ line 43) in openpoints/loss/bulid.py.