siamese-fc: A Matlab repository from wjpeters

Fully-Convolutional Siamese Networks for Object Tracking

The code in this repository enables you to reproduce the experiments of our paper. It can be used in two ways: (1) tracking only and (2) training and tracking.

Project page: http://www.robots.ox.ac.uk/~luca/siamese-fc.html

If you find our work and/or curated dataset useful, please cite:

@inproceedings{bertinetto2016fully,
  title={Fully-Convolutional Siamese Networks for Object Tracking},
  author={Bertinetto, Luca and Valmadre, Jack and Henriques, Jo{\~a}o F and Vedaldi, Andrea and Torr, Philip H S},
  booktitle={ECCV 2016 Workshops},
  pages={850--865},
  year={2016}
}

[ Tracking only ] If you don't care much about training, simply plug one of our pretrained networks to our basic tracker and see it in action.

Prerequisites: GPU, CUDA drivers, cuDNN, Matlab (we used 2015b), MatConvNet (we used v1.0-beta20).
Clone the repository.
Download one of the pretrained networks from http://www.robots.ox.ac.uk/~luca/siamese-fc.html
Go to siam-fc/tracking/ and remove the trailing .example from env_paths_tracking.m.example, startup.m.example and run_tracking.m.example, editing the files as appropriate.
Be sure to have at least one video sequence in the appropriate format. You can find an example here in the repository (siam-fc/demo-sequences/vot15_bag).
siam-fc/tracking/run_tracking.m is the entry point to execute the tracker, have fun!

[ Training and tracking ] Well, if you prefer to train your own network, the process is slightly more involved (but also more fun).

Prerequisites: GPU, CUDA drivers, cuDNN, Matlab (we used 2015b), MatConvNet (we used v1.0-beta20).
Clone the repository.
Follow these step-by-step instructions, which will help you generating a curated dataset compatible with the rest of the code.
If you did not generate your own, download the imdb_video.mat (6.7GB) with all the metadata and the dataset stats.
Go to siam-fc/training/ and remove the trailing .example from env_paths.m.example, startup.m.example and run_experiment.m.example editing the files as appropriate.
siam-fc/training/run_experiment.m is the entry point to start training. Default hyper-params are at the start of experiment.m and can be overwritten by custom ones specified in run_experiment.m.
By default, training plots are saved in siam-fc/training/data/. When you are happy, grab a network snapshot (net-epoch-X.mat) and save it somewhere convenient to use it for tracking.
Go to point 4. of Tracking only and enjoy the result of the labour of your own GPUs!

wjpeters/siamese-fc

Fully-Convolutional Siamese Networks for Object Tracking