Improved Dense Trajectory features encoded by Fisher Vectors
This is the implementations of DTF+Fisher vector on UCF101 dataset. VGG fisher vector implementation(http://www.robots.ox.ac.uk/~vgg/software/enceval_toolkit/.) and VLFfeat(http://www.vlfeat.org/) package are used. LIBSVM is also needed for one-vs-rest classification.
For theoretical introduction about this project, please refer to posts Action Recognition with Fisher Vectors and DTF Notes on my blog.
To run the program, please execute master in Matlab.
[Warning for Windows users]: Shell script gzip_dtf_files is used to compress DTF data in compute_fisher.m and subsample.m. If you don't have Shell installed, please remove the related lines.