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.