Associatively Segmenting Instances and Semantics in Point Clouds

The full paper is available at: https://arxiv.org/abs/1902.09852. Qualitative results of ASIS on the S3DIS and vKITTI test fold:

Overview

Dependencies

The code has been tested with Python 2.7 on Ubuntu 14.04.

TensorFlow
h5py

Data and Model

Download 3D indoor parsing dataset (S3DIS Dataset). Version 1.2 of the dataset is used in this work.

python collect_indoor3d_data.py
python gen_h5.py
cd data && python generate_input_list.py
cd ..

(optional) Trained model can be downloaded from here.

Usage

Compile TF Operators

Refer to PointNet++
Training

cd models/ASIS/
ln -s ../../data .
sh +x train.sh 5

Evaluation

python eval_iou_accuracy.py

Note: We test on Area5 and train on the rest folds in default. 6 fold CV can be conducted in a similar way.

Citation

If our work is useful for your research, please consider citing:

@inproceedings{wang2019asis,
	title={Associatively Segmenting Instances and Semantics in Point Clouds},
	author={Wang, Xinlong and Liu, Shu and Shen, Xiaoyong and Shen, Chunhua, and Jia, Jiaya},
	booktitle={CVPR},
	year={2019}
}

Acknowledgemets

This code largely benefits from following repositories: PointNet++, SGPN, DGCNN and DiscLoss-tf