This project implements ASAP-Net of paper ASAP-Net: Attention and Structure Aware Point Cloud Sequence Segmentation (BMVC2020).
We improve spatio-temporal point cloud feature learning with a flexible module called ASAP module considering both attention and structure information across frames, which can be combined with different backbones. Incorporating our module into backbones brings semantic segmentation performance improvements on both Synthia and SemanticKITTI datasets (+3.4 to +15.2 mIoU points with different backbones).
The Synthia experiments is implemented with TensorFlow and the SemanticKITTI experiments is implemented with PyTorch. We tested the codes under TensorFlow 1.13.1 GPU version, PyTorch 1.1.0, CUDA 10.0, g++ 5.4.0 and Python 3.6.9 on Ubuntu 16.04.12 with TITAN RTX GPU. For SemanticKITTI experiments, you should have a GPU memory of at least 16GB.
We use the implementation in xingyul/meteornet. Please follow the instructions below.
The TF operators are included under Synthia_experiments/tf_ops
, you need to compile them first by make
under each ops subfolder (check Makefile
) or directly use the following commands:
cd Synthia_experiments
sh command_make.sh
Please update arch
in the Makefiles for different CUDA Compute Capability that suits your GPU if necessary.
We use the PoinNet++ implementation in sshaoshuai/Pointnet2.PyTorch. Use the commands below to build Torch operators.
cd SemanticKITTI_experiments/ASAP-Net_PointNet2/pointnet2
python setup.py install
The codes for experiments on Synthia is in Synthia_experiments/semantic_seg_synthia
. Please refer to Synthia_experiments/semantic_seg_synthia/README.md
for more information on data preprocessing and running instructions.
The SemanticKITTI_experiments/ImageSet2 folder contains dataset split information. Please put it under your semanticKITTI dataset like Path to semanticKITTI dataset/dataset/sequences
.
The codes for framework with PointNet++ as Backbone is in SemanticKITTI_experiments/ASAP-Net_PointNet2
. Please refer to SemanticKITTI_experiments/ASAP-Net_PointNet2/README.md
for more information on running instructions.
The codes for framework with SqueezeSegV2 as Backbone is in SemanticKITTI_experiments/ASAP-Net_SqueezeSegV2
. Please refer to SemanticKITTI_experiments/ASAP-Net_SqueezeSegV2/README.md
for more information on running instructions.
Special thanks for open source codes including xingyul/meteornet, sshaoshuai/Pointnet2.PyTorch and PRBonn/lidar-bonnetal.
Please cite these papers in your publications if it helps your research:
@article{caoasap,
title={ASAP-Net: Attention and Structure Aware Point Cloud Sequence Segmentation},
author={Cao, Hanwen and Lu, Yongyi and Lu, Cewu and Pang, Bo and Liu, Gongshen and Yuille, Alan}
booktitle={British Machine Vision Conference (BMVC)},
year={2020}
}