BichenWuUCB/SqueezeSeg

SqueezeSeg: Convolutional Neural Nets with Recurrent CRF for Real-Time Road-Object Segmentation from 3D LiDAR Point Cloud

By Bichen Wu, Alvin Wan, Xiangyu Yue, Kurt Keutzer (UC Berkeley)

This repository contains a tensorflow implementation of SqueezeSeg, a convolutional neural network model for LiDAR segmentation. A demonstration of SqueezeSeg can be found below:

Please refer to our video for a high level introduction of this work: https://youtu.be/Xyn5Zd3lm6s. For more details, please refer to our paper: https://arxiv.org/abs/1710.07368. If you find this work useful for your research, please consider citing:

@article{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},
    journal={ICRA},
    year={2018}
}
@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{yue2018lidar,
    title={A lidar point cloud generator: from a virtual world to autonomous driving},
    author={Yue, Xiangyu and Wu, Bichen and Seshia, Sanjit A and Keutzer, Kurt and Sangiovanni-Vincentelli, Alberto L},
    booktitle={ICMR},
    pages={458--464},
    year={2018},
    organization={ACM}
}

We recently open-sourced the code for SqueezeSegV2, a follow-up work to SqueezeSeg with significantly improved performance. For details, please check out: https://github.com/xuanyuzhou98/SqueezeSegV2

License

SqueezeSeg is released under the BSD license (See LICENSE for details). The dataset used for training, evaluation, and demostration of SqueezeSeg is modified from KITTI raw dataset. For your convenience, we provide links to download the converted dataset, which is distrubited under the Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License.

Installation:

The instructions are tested on Ubuntu 16.04 with python 2.7 and tensorflow 1.0 with GPU support.

Clone the SqueezeSeg repository:
```
git clone https://github.com/BichenWuUCB/SqueezeSeg.git
```
We name the root directory as $SQSG_ROOT.
Setup virtual environment:
1. By default we use Python2.7. Create the virtual environment
```
virtualenv env
```
2. Activate the virtual environment
```
source env/bin/activate
```
Use pip to install required Python packages:
```
pip install -r requirements.txt
```

Demo:

To run the demo script:
```
cd $SQSG_ROOT/
python ./src/demo.py
```
If the installation is correct, the detector should write the detection results as well as 2D label maps to $SQSG_ROOT/data/samples_out. Here are examples of the output label map overlaped with the projected LiDAR signal. Green masks indicate clusters corresponding to cars and blue masks indicate cyclists.

Training/Validation

First, download training and validation data (3.9 GB) from this link. This dataset contains LiDAR point-cloud projected to a 2D spherical surface. Refer to our paper for details of the data conversion procedure. This dataset is converted from KITTI raw dataset and is distrubited under the Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License.
```
cd $SQSG_ROOT/data/
wget https://www.dropbox.com/s/pnzgcitvppmwfuf/lidar_2d.tgz
tar -xzvf lidar_2d.tgz
rm lidar_2d.tgz
```
Now we can start training by
```
cd $SQSG_ROOT/
./scripts/train.sh -gpu 0 -image_set train -log_dir ./log/
```
Training logs and model checkpoints will be saved in the log directory.

We can launch evaluation script simutaneously with training

cd $SQSG_ROOT/
./scripts/eval.sh -gpu 1 -image_set val -log_dir ./log/

We can monitor the training process using tensorboard.
```
tensorboard --logdir=$SQSG_ROOT/log/
```
Tensorboard displays information such as training loss, evaluation accuracy, visualization of detection results in the training process, which are helpful for debugging and tunning models, as shown below: