chensun11 / dtfv

Fisher Vector implementation for Dense Trajectory features

Fisher Vector implementation for Dense Trajectories (DTFV)

This is a C++ implementation of Fisher Vector (FV) for Dense Trajectory (DT) features.

DTFV provides a binary with a Python script to generate Fisher Vectors. It also provides code to sample feature points, and to train PCA and GMM codebooks.

Some highlights include:

• Provides a full pipeline from DT raw feature to Fisher Vectors
• Aggregate feature points on-the-fly, no need to load all points in memory
• Raw features are piped to the Fisher Vector generator, no need to store the large raw features
• State-of-the-art performance for large action/event video datasets

It depends on:

• Dense Trajectories (DT)

  To extract DT features

• VLFeat

  To train Gaussian Mixture Models

• ALGLIB

  To train PCA projection matrix

Using linear SVM, we tested the code on two action/event classification datasets:

• TRECVID Multimedia Event Detection (MED)

  Trained on EX100, tested on MEDTest. The mean average precision for event 6 to 15 and 21 to 30 is 0.33.

• UCF 101

  Partitions were defined by THUMOS workshop. The average classification accuracy is 85%.

The pretrained codebooks on these two dataset are also provided.

Please see our WACV 2013 paper for more details.

dtfv's Issues

How to train the codebook on my own data?

Could you please give some guide on training the codebook on my own dataset?

Thanks.

The fv results is always nan?

I have used your dtfv code to do fisher encoding. At first, I used it on my own dataset by rerunning exract_fv.py. But the results in outputDir is always nan,for some files , they are all 0.
When I test the code on UCF101, the results remained nan. I don't know what went wrong.Do you have any ideas?Thanks a lot !

what is the function of alglib?

hello, i want to operation this program under Windows. I don't know the function of src/alglib ,can you give me some advice? how to add the src/alglib to a vs program?