Attention-based relation and context modeling for point cloud semantic segmentation (SMI 2020)

Zhiyu Hu, Dongbo Zhang, Shuai Li, Hong Qin

Overview

Environment

Our method is tested in Python 2.7 and TensorFlow 1.4.0 on a workstation with an Intel Core i7-4790 CPU (3.60GHz, 16GB memory) and a GeForce GTX 1070 GPU (8GB memory, CUDA 8.0).

Data

• Download S3DIS Dataset. Version 1.2 (v1.2_Aligned_Version) 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 ..

Usage

• Create Conda Environment

  conda create -n ARCM python=2.7
  source activate ARCM

• Clone the Repository

git clone https://github.com/hu-zhiyu/ARCM && cd ARCM

• Install Required Dependencies

pip install -r requirements.txt

• Compile TF Operators

cd tf_ops
sh tf_compile_all.sh

Make sure no errors generated at this step and refer to PointNet++ when encountering any problems.

• Test with Pre-trained Model

cd models/
python test.py --gpu 0 --log_dir log5_test --model_path log5/epoch_99.ckpt --input_list  meta/area5_data_label.txt --verbose
python eval_iou_accuracy.py

S3DIS Area 5 Quantitative Results

OA	mAcc	mIoU	ceiling	floor	wall	beam	column	window	door	table	chair	sofa	bookcase	board	clutter
88.50	66.09	59.66	93.38	98.45	81.50	0.00	7.00	55.14	48.61	77.16	87.81	50.68	65.54	57.76	52.57

S3DIS Visual Results

Acknowledgements

This code largely benefits from following repositories:

• PointNet++
• ASIS