Video summarization based on Deep Semantic Features

The implementation is based on the paper "Video Summarization using Deep Semantic Features" in ACCV'16 [arXiv]

Install dependency

The source is running on Python 2.7. Anaconda and Jupyter notebook are utilized for creating the ideal environment Download. The environment can be installed via conda environment:

$ conda create --name <env> --file <environment file>

For example:

$ conda create --name video_summarization --file requirements.txt

This code utilizes tools provided by M. Gygli et al. [1]. You can set it up by:

cd vsum_dsf
git clone https://github.com/gyglim/gm_submodular.git
cd gm_submodular
python setup.py install --user

[1] Gygli, Grabner & Van Gool. Video Summarization by Learning Submodular Mixtures of Objectives. CVPR 2015.

* The source are already set up with gm_submodular.

Download dataset and model parameters

The source can be tested on Summe dataset provided bt M. Gygli et al. [2]. To test the model in the paper, download a data.zip HERE and extract it in the folder vsum_dsf.

[2] Gygli M., Grabner, H., Riemenschneider, H., van Gool, L.: Creating summaries from user videos. ECCV 2014.

* Data are already contained in the source code

Script

See the Demo.ipynb or run the script below to generate the video summaries:

$ python script/summarize.py

For evaluation:

$ python script/evaluate.py results/summe/smt_feat