End-to-End Trainable Fully-Convolutional Siamese Networks for Video Object Segmentation with Bounding Box

This is an implementation of SiamVOS (My Thesis: End-to-End Trainable Fully-Convolutional Siamese Networks for Video Object Segmentation with Bounding Box) in pytorch.

Compared to conventional semi-supervised video object segmentation methods,
SiamVOS requires only a bounding box of the target for video object segmentation.

Environment setup

All the code has been tested on Ubuntu18.04, python3.6, Pytorch1.2.0, anaconda2019.03, CUDA 10.0, GTX TITAN x GPU

• Clone the repository

  git clone https://github.com/nijkah/SiamVOS.git && cd SiamVOS

• Setup python environment

  conda create -n SiamVOS python=3.6
  conda activate SiamVOS
  conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=10.0 -c pytorch
  pip install -r requirements.txt

  NOTICE imgaug library has an issue on latest numpy=1.8.1,
  so pip install git+https://github.com/aleju/imgaug rather than pip install imgaug
  For more information see this

• Download data

DAVIS-2017 Youtube-VOS GyGO

and set paths in cfg.py

and you can download the pre-trained deeplab model from
here.
Put this in the data folder.

• train the model

  cd scripts
  python train_siamvos.py

• evaluate the model

  python evaluate_siamvos.py

  You can download the trained SiamVOS model from
  here.
  Put this in the data/snapshots folder.

Results