Less Is More: Label Recommendation for Weakly Supervised Point Cloud Semantic Segmentation

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

Installation Guide

Prerequisites

Before installing our program, you should install PointNeXt.

Installation steps

1. Move this project's files to the root directory of PointNeXt.
2. Replace the corresponding files according to the guide.

Usages

Dataset The presampling collects all point clouds, area by area and room by room, following PointNeXt. You can download our preprocessed S3DIS dataset as follows:

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

Inductive Bias Learning and Recommendation Please modify the corresponding configuration files to use your own file path.

For example, train PointNeXt++ with point cloud upsampling as pretext task, and then recommend with single-scene clustering strategy (kmeans, the code is based on ContrastiveSceneContexts) as follows

CUDA_VISIBLE_DEVICES=0 python examples/segmentation/main.py --cfg cfgs/s3dis_LiM/pointnet++_upsampling.yaml visualize=True

and using a cross-scene clustering strategy as follows

CUDA_VISIBLE_DEVICES=0 python examples/segmentation/main.py --cfg cfgs/s3dis_LiM/pointnet++_upsampling_batch.yaml visualize=True

Point Cloud Semantic Segmentation Learning

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