Video Features Extractor

Extracting several kinds of visual representations from videos.

Supported visual features

The following frame-level (*_features) and video-level (*_gloabal) visual representations are supported:

1. 2D-CNN (cnn_features, cnn_globals, cnn_sem_globals)
2. 3D-CNN (c3d_features, c3d_globals, i3d_features, i3d_globals)
3. ECO (eco_features, eco_globals, eco_sem_features, eco_sem_globals)
4. TSM (tsm_features, tsm_globals, tsm_sem_features, tsm_sem_globals)

Note: *_sem_* representations are based on the classification level (probability distribution) of respective models.

Supported Datasets

This package has been tested for extracting visual representations from videos of the following video-caption datasets:

1. MSVD
2. M-VAD
3. MSR-VTT
4. TRECVID-2020
5. TRECVID-2020-Test
6. TGIF
7. VATEX
8. ActivityNet
9. ActivityNet-Test
10. ActivityNet-Fragments