Long and untrimmed video learning has recieved increasing attention in recent years. This repo aims to provide some simple and effective scripts for long and untrimmed video feature extraction. We adopt the video processing pipline from TSP and adapt it with several awesome vision pretraining backbones.

Run conda env create -f base_environment.yaml for base environment setup. For specific model setup, please check their project link:

Video Swin Transformer

Omnivore

CLIP

Run bash Scripts/generate_video_metada.sh to extract metadata from video, where VIDEO_FOLDER is the directory contains the raw videos, and OUTPUT_CSV_PATH is the output csv file contains the generated video metadata.

Then run the following script to extract features:

bash Scripts/extract_${MODEL_NAME}_feat.sh

Before running, rember to set the defined variable in the script.

Finally, run bash Scripts/merge_pkl_to_h5.sh to merge the video features to a single h5 file.

This repo is mainly based on pipeline provided by TSP.