Algorithm used for unsupervised action tube extraction from videos. The method is described here.
This repository contains a Matlab implementation of the code, and has been tested on Linux, using Matlab R2015a.
If you find this unsupervised tube extraction algorithm useful in your research, please consider citing:
@inproceedings{marian2015unsupervised,
title={Unsupervised Tube Extraction Using Transductive Learning and Dense Trajectories},
author={Marian Puscas, Mihai and Sangineto, Enver and Culibrk, Dubravko and Sebe, Nicu},
booktitle={Proceedings of the IEEE International Conference on Computer Vision},
pages={1653--1661},
year={2015}
}
The algorithm is under the MIT License, details in LICENSE
- extract the absolute coordinates of the trajectories throughout the video using Improved Dense Trajectories with default parameters.
- to maintain a roughly constant number of trajectories in the last frames of the video, we have mirrored the last 3 frames.
- for cnn feature extraction use the bvlc_reference_caffenet model - fc7 features