nijkah / siamvos

Video Object Segmentation requiring only a bounding box

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

Model	DAVIS2016 mean IoU	DAVIS2017
SiamMask (paper)	71.7	51.1
SiamVOS	74.3	53.2