Home

Awesome

SqueezeSegV3: Spatially-Adaptive Convolution for Efficient Point-Cloud Segmentation. (ECCV2020)

By Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka.

This repository contains a Pytorch implementation of SqueezeSegV3, a state-of-the-art model for LiDAR segmentation. The framework of our SqueezeSegV3 can be found below:

<p align="center"> <img src="./figure/framework.png"/ width="750"> </p>

Selected quantitative results of different approaches on the SemanticKITTI dataset (* means KNN post-processing):

MethodmIoUcarbicyclemotorcycletruckpersonbicyclistmotorcyclistroad
SqueezeSeg29.568.816.04.13.312.913.10.985.4
SqueezeSegV239.781.818.517.913.420.125.13.988.6
RangeNet2147.485.426.226.518.631.833.64.091.4
RangeNet5349.986.424.532.725.536.233.64.791.8
SqueezeSegV3-2148.884.631.532.411.339.436.121.390.8
SqueezeSegV3-5352.987.435.233.729.041.839.120.191.8
SqueezeSegV3-21*51.689.433.734.911.342.644.921.290.8
SqueezeSegV3-53*55.992.538.736.529.645.646.220.191.7

Visualization results of SqueezeSegV3:

<p align="center"> <img src="./figure/sample3.gif" width="425" height="450" /> <img src="./figure/sample4.gif" width="425" height="450" /> </p>

For more details, please refer to our paper: SqueezeSegV3. The work is a follow-up work to SqueezeSeg, SqueezeSegV2 and LATTE. If you find this work useful for your research, please consider citing:

@inproceedings{xu2020squeezesegv3,
  title={Squeezesegv3: Spatially-adaptive convolution for efficient point-cloud segmentation},
  author={Xu, Chenfeng and Wu, Bichen and Wang, Zining and Zhan, Wei and Vajda, Peter and Keutzer, Kurt and Tomizuka, Masayoshi},
  booktitle={European Conference on Computer Vision},
  pages={1--19},
  year={2020},
  organization={Springer}
}

Related works:

 @inproceedings{wu2018squeezesegv2,
   title={SqueezeSegV2: Improved Model Structure and Unsupervised Domain Adaptation
   for Road-Object Segmentation from a LiDAR Point Cloud},
   author={Wu, Bichen and Zhou, Xuanyu and Zhao, Sicheng and Yue, Xiangyu and Keutzer, Kurt},
   booktitle={ICRA},
   year={2019},
 }
 
@inproceedings{wu2017squeezeseg,
   title={Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud},
   author={Wu, Bichen and Wan, Alvin and Yue, Xiangyu and Keutzer, Kurt},
   booktitle={ICRA}, 
   year={2018}
 }
 
@inproceedings{wang2019latte,
  title={LATTE: accelerating lidar point cloud annotation via sensor fusion, one-click annotation, and tracking},
  author={Wang, Bernie and Wu, Virginia and Wu, Bichen and Keutzer, Kurt},
  booktitle={2019 IEEE Intelligent Transportation Systems Conference (ITSC)},
  pages={265--272},
  year={2019},
  organization={IEEE}
}

License

SqueezeSegV3 is released under the BSD license (See LICENSE for details).

Installation

The instructions are tested on Ubuntu 16.04 with python 3.6 and Pytorch 1.1.0 with GPU support.

git clone https://github.com/chenfengxu714/SqueezeSegV3.git
pip install -r requirements.txt

Pre-trained Models

The pre-trained SqueezezSegV3-21 and SqueezeSegV3-53 are avaliable at Google Drive, you can directly download the two files.

Demo

We provide a demo script:

cd ./src/tasks/semantic/
python demo.py -m /path/to/model

You can find the prediction .label files and projected map in ./src/sample_output file, an example is shown below:

<p align="center"> <img src="./figure/000000.png"/ width="800"> </p>

Inference

To infer the predictions for the entire dataset:

cd ./src/tasks/semantic/
python infer.py -d /path/to/dataset/ -l /path/for/predictions -m /path/to/model

To visualize the prediction for the sequence point cloud:

python visualize.py -d /path/to/dataset/ -p /path/to/predictions/ -s SQ_Number

Training

cd ./src/tasks/semantic/

To train a network (from scratch):

python train.py -d /path/to/dataset -ac /config/arch/CHOICE.yaml -l /path/to/log

To train a network (from pretrained model):

python train.py -d /path/to/dataset -ac /config/arch/CHOICE.yaml -l /path/to/log -p /path/to/pretrained

We can monitor the training process using tensorboard.

tensorboard --logdir /file_path/

Evaluation

python evaluate_iou.py -d /path/to/dataset -p /path/to/predictions/ --split valid

Credits

We referred to RangeNet++ (Paper, Code) during our development. We thank the authors of RangeNet++ for open-sourcing their code.