This is the implementation code of the paper "Conditional Video Summarization via Non-monotone Submodular Function Maximization".

Get started:

1. System requirement: Pytorch 1.1.0, Python 3.7
2. Download the datasets from the Google drive link and place all datasets under the "datasets" folder. It is about 1 GB and will take some time to finish.

Generic Video Summarization:

Use OVP dataset

1. GoogleNet features
   python QVSmain.py -d datasets/qvs_dataset_ovp_google_pool5.h5 -s datasets/ovpsplitsfull.json -m OVP --mode 2
2. Histgram color features
   python QVSmain.py -d datasets/qvs_dataset_ovp_color.h5 -s datasets/ovpsplitscolor.json -m OVP --mode 2 --ftype color

Use Youtube dataset

1. GoogleNet features
   python QVSmain.py -d datasets/qvs_dataset_youtube_google_pool5.h5 -s datasets/youtubesplitsfull.json -m Youtube --mode 2
2. Histgram color features
   python QVSmain.py -d datasets/qvs_dataset_youtube_color.h5 -s datasets/youtubesplitscolor.json -m Youtube --mode 2 --ftype color

Conditional Video Summarization:

Using Youtube v73 as an example

1. Focus on the game field
   python QVSmain.py -d datasets/qvs_dataset_youtube_google_pool5.h5 -s datasets/youtubesplitsv73.json -m Youtube --mode 2 --cond --query 2
2. Focus on the stands
   python QVSmain.py -d datasets/qvs_dataset_youtube_google_pool5.h5 -s datasets/youtubesplitsv73.json -m Youtube --mode 2 --cond --query 9