CycleSiam

This is the official implementation code for CycleSiam -- "Self-supervised Object Tracking and Segmentation with Cycle-consistent Siamese Networks". It is built based on SiamMask. For technical details, please refer to :

Self-supervised Object Tracking and Segmentation with Cycle-consistent Siamese Networks
Weihao Yuan, Michael Yu Wang, Qifeng Chen
IROS2020
[Paper]

Bibtex

If you find this code useful, please consider citing:

@inproceedings{yuan2020self,
  title={Self-supervised object tracking and segmentation with cycle-consistent siamese networks},
  author={Yuan, Weihao and Wang, Michael Yu and Chen, Qifeng},
  booktitle={Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages={},
  year={2020},
  organization={IEEE}
}

Contents

1. Environment Setup
2. Demo
3. Training

Environment setup

This code has been tested on Ubuntu 16.04, Python 3.6, Pytorch 0.4.1, CUDA 9.2, RTX 2080 GPUs

• Clone the repository

git clone https://github.com/weihaosky/CycleSiam.git && cd CycleSiam
export CycleSiam=$PWD

• Setup python environment

conda create -n cyclesiam python=3.6
source activate cyclesiam
pip install -r requirements.txt
bash make.sh

• Add the project to your PYTHONPATH

export PYTHONPATH=$PWD:$PYTHONPATH

Demo

• Setup your environment
• Download the CycleSiam model

cd $CycleSiam/experiments/siammask_sharp

• Run demo.py

cd $CycleSiam/experiments/siammask_sharp
export PYTHONPATH=$PWD:$PYTHONPATH
python ../../tools/demo.py --resume checkpoint_cyclesiam_plus.pth --config config_davis.json

Training

Training Data

• Download the Youtube-VOS, COCO, ImageNet-DET, and ImageNet-VID.
• Preprocess each datasets according the readme files.

Download the pre-trained model (174 MB)

(This model was trained on the ImageNet-1k Dataset)

cd $CycleSiam/experiments
wget http://www.robots.ox.ac.uk/~qwang/resnet.model
ls | grep siam | xargs -I {} cp resnet.model {}

Training CycleSiam base model

• Setup your environment
• From the experiment directory, run

cd $CycleSiam/experiments/siammask_base/
bash run.sh

• If you experience out-of-memory errors, you can reduce the batch size in run.sh.
• You can view progress on Tensorboard (logs are at <experiment_dir>/logs/)
• After training, you can test checkpoints on VOT dataset.

bash test_all.sh -s 1 -e 20 -d VOT2018 -g "0 1 2 3"  # test all snapshots with 4 GPUs

• Select best model for hyperparametric search.

#bash test_all.sh -m [best_test_model] -d VOT2018 -n [thread_num] -g [gpu_num] # 8 threads with 4 GPUS
bash test_all.sh -m snapshot/checkpoint_e18.pth -d VOT2018 -n 8 -g "0 1 2 3" # 8 threads with 4 GPUS

Training CycleSiam model with the Refine module

• Setup your environment
• In the experiment file, train with the best CycleSiam base model

cd $CycleSiam/experiments/siammask_sharp
bash run.sh <best_base_model>
bash run.sh checkpoint_e18.pth

• You can view progress on Tensorboard (logs are at <experiment_dir>/logs/)
• After training, you can test checkpoints on VOT dataset

bash test_all.sh -s 1 -e 20 -d VOT2018 -g "0 1 2 3"

• Select best model for hyperparametric search.

#bash test_all.sh -m [best_test_model] -d VOT2018 -n [thread_num] -g [gpu_num] # 8 threads with 4 GPUS
bash test_all.sh -m snapshot/checkpoint_e19.pth -d VOT2018 -n 8 -g "0 1 2 3" # 8 threads with 4 GPUS

Pretrained models

License

Licensed under an MIT license.