TVRetrieval/utils/video_feature at master · jayleicn/TVRetrieval · GitHub

Name		Name	Last commit message	Last commit date
parent directory ..
README.md		README.md
convert_feature_frm_to_clip.py		convert_feature_frm_to_clip.py
convert_feature_frm_to_clip.sh		convert_feature_frm_to_clip.sh
extract_i3d_features.py		extract_i3d_features.py
extract_i3d_features.sh		extract_i3d_features.sh
extract_image_features.py		extract_image_features.py
extract_resnet152_2048_features.sh		extract_resnet152_2048_features.sh
i3d.py		i3d.py
merge_align_i3d.py		merge_align_i3d.py
merge_align_i3d.sh		merge_align_i3d.sh
normalize_and_concat.py		normalize_and_concat.py
normalize_and_concat.sh		normalize_and_concat.sh

README.md

video feature extraction

I3D feature extraction requirements:

tensorflow-gpu==1.14
dm-sonnet-gpu==1.32
opencv-python

ResNet-152 feature extraction requirements:

PyTorch
Torchvision

Note the video features released at tvr_feature_release.tar.gz is extracted from 15 FPS frames, which is not publicly available (we only released 3 FPS frames).