A PyTorch reimplementation of FCSN in paper "Video Summarization Using Fully Convolutional Sequence Networks".
Mrigank Rochan, Linwei Ye, and Yang Wang.
Video Summarization Using Fully Convolutional Networks.
European Conference on Computer Vision (ECCV), 2018
A TVSum dataset (downsampled to 320 frames per video) preprocessed by make_dataset.py is available here. There are 50 groups in this hdf5 file named video_1 to video_50 . Datasets in each group is as follows:
| name | description |
|---|---|
length |
scalar, number of video frames |
feature |
shape (320, 1024) |
label |
shape (320, ) |
change_points |
shape (n_segments, 2) stores begin and end of each segment |
n_frame_per_seg |
shape (n_segments, ) number of frames in each segment |
user_summary |
shape (20, length) summary from 20 users, each row is a binary vector |
First change the data_path in config.py to your own hdf5 data path. Then run
python train.pyEvery 5 epoch, model parameters and predicted summaries will be saved in save_dir and score_dir respectively. Evaluation result (precision, recall and f-score) will be printed at the same time.
To generate keyframes (images) and keyshots (video), run
python gen_summary.py --h5_path {your hdf5 dataset path} --json_path {json file path saved in score_dir} --data_root {root dir of tvsum dataset} --save_dir {where to save the generating summary} --barpreview of score-bar (x-axis: frame index, y-axis: user's score, and columns in orange is selected keyshots):
