One-Shot Video Object Segmentation (OSVOS)

Visit our project page for accessing the paper, and the pre-computed results.

This is the implementation of our work One-Shot Video Object Segmentation (OSVOS), for semi-supervised video object segmentation. OSVOS is based on a fully convolutional neural network architecture that is able to successively transfer generic semantic information, learned on ImageNet, to the task of foreground segmentation, and finally to learning the appearance of a single annotated object of the test sequence (hence one-shot). Experiments on DAVIS 2016 show that OSVOS is faster than currently available techniques and improves the state of the art by a significant margin (79.8% vs 68.0%).

While the results of the paper were obtained by this code, we also provide a TensorFlow implementation of OSVOS: OSVOS-TensorFlow.

NEW: PyTorch implementation also available: OSVOS-PyTorch!

Installation:

Clone the OSVOS-caffe repository

git clone https://github.com/kmaninis/OSVOS-caffe.git

Install the Caffe version under caffe-osvos/ along with standard dependencies, pycaffe and matcaffe. Caffe would need to be built with support for Python layers, in case you would like to use the Python API (TODO). cuDNN is not necessary.
```
# In your Makefile.config, make sure to have this line uncommented
WITH_PYTHON_LAYER := 1
```
Download the parent model from here (55 MB) and put it under models/.
Optionally download the contour model for contour snapping from here (55 MB) and put it under models/.
If you want to use the contour snapping step (a.k.a you downloaded the model of step 4.), run build.m from within MATLAB.
All the steps to re-train OSVOS are provided in this repository. In case you would like to test with the pre-trained models, you can download them from here (1GB) and put it under models/.

Demo online training and testing

Edit in file set_params.m the parameters of the code (eg. useGPU, gpu_id, etc.).
Run demo.m.
You can test all sequences of DAVIS 2016 validation set, by running test_all.m, once the pre-trained models are available under models/.

It is possible to work with all sequences of DAVIS 2016 just by creating a soft link (ln -s /path/to/DAVIS/) in the root folder of the project.

Training the parent network (optional)

All necessary files are under src/parent. So, cd src/parent.
Download the pre-trained vgg model by running ./download_pretrained_vgg.sh
Augment the data. In the paper we used flipping and scaling into 0.5, 0.8 and 1.0 of the original scale. Your image and ground truth pairs are specified in solvers/train_pair.txt.
Under solvers edit the data_root_dir of train_val*.prototxt.
Finally, train the parent model with python solve_cluster.py. You need pycaffe for this step, so don't forget to make pycaffe when installing Caffe.

Enjoy! :)

Citation

If you use this code, please consider citing the following paper:

@Inproceedings{Cae+17,
  Title          = {One-Shot Video Object Segmentation},
  Author         = {S. Caelles and K.K. Maninis and J. Pont-Tuset and L. Leal-Taix\'e and D. Cremers and L. {Van Gool}},
  Booktitle      = {Computer Vision and Pattern Recognition (CVPR)},
  Year           = {2017}
}

If you encounter any problems with the code, want to report bugs, etc. please contact me at kmaninis[at]vision[dot]ee[dot]ethz[dot]ch.

wuzzh/OSVOS-caffe