This repository is intended for generating video summaries using 4 cutting edge summarization models: PGL-SUM, CA-SUM, DSNet anchor based and DSNet anchor free. Models are pretrained on TVSum and SumMe datasets.
Create virtual environment:
python -m venv .summarization
Activate virtual environment:
source .summarization/bin/activate
Install dependencies:
pip install -r requirements.txt
Summary generation:
First move to the source folder:
cd src
Single video:
python inference.py pglsum --source ../custom_data/videos/source_video_name.mp4 --save-path ./output/summary_video_name.mp4 --sample-rate 30 --final-frame-length 30
Folder of videos:
python inference.py pglsum --source ../custom_data/videos/source_video_folder --save-path ./output/summary_videos_folder --sample-rate 30 --final-frame-length 30
Eligible model names: pglsum - PGL-SUM, casum - CA-SUM, dsnet_ab - DSNet anchor based and dsnet_af - DSNet anchor free.
--sample-rate 30 means the model will take every 30th frame for analysis
--final-frame-length 27 means the final video summary will have 30 frames (around 27 sec video, 23-31 sec depending on frames per second in the initial video)
--max-shot-length 8 means a single shot won't be longer than 8 frames
--min-penalty-shot-length 5 means that a shot of 5 or less frames will have length penalty and therefore will be less likely to appear in the final summary