Evaluating and Extending Unsupervised VideoSummarization Methods

1. Evaluation unsupervised methods with different metric under same configuration.
2. Investigating effect of extracted features in unsupervised methods and extend it to perform better than baseline work.
3. Identify gaps form last step and try to fill by porposed solution by extending or modifying an existing model.

Abstract

This study validates the recent work in unsupervised video summarization and extends the experiments with feature variation. In our understanding, deep learning based approaches select frame-candidates for a video summary. These frames are drawn with a probability that can be used to calculate the scaled importance for comparison with the basic truth as an evaluation. However, most works in the literature focus on network architecture and follow the same feature extraction technique using the resulting deep features from one pretrained model. At this point a notable gap of the existing approaches is feature variation, which this study explores.
Building on that, model variables like network architecture, optimizer, and activation functions can have an impact on the performance combined with different feature selection techniques that haven’t beenexplored yet.Further, the Evaluation of existing models is conducted using one metric,which may be not representative in video summarization task, since it could ignore key-frames in favor of other frames and still have a relatively high value. This work performs feature extraction using multiple pretrained neural models, and then measures the impact of them on current state-of-the-art works.Then,it evaluates the state-of-the-art works using different evaluation metrics. Eventually, it aims to find an unsupervised video summarization method, in order to fill the gaps, and leverage the existing works.

Important Wiki Pages:

• Notes on SUM-GAN-AAE (Apostolidis et al. 2020)

• References and important findings

• Reproduce (Zhou et al. 2018)

• Conda cheat sheet

Datasets

Structured h5 files with the video features and annotations of the SumMe and TVSum datasets are available within the "data" folder. The GoogleNet features of the video frames were extracted by Ke Zhang and [Wei-Lun Chao] and the h5 files were obtained from Kaiyang Zhou.

These files have the following structure:

/key
    /features                 2D-array with shape (n_steps, feature-dimension)
    /gtscore                  1D-array with shape (n_steps), stores ground truth improtance score (used for training, e.g. regression loss)
    /user_summary             2D-array with shape (num_users, n_frames), each row is a binary vector (used for test)
    /change_points            2D-array with shape (num_segments, 2), each row stores indices of a segment
    /n_frame_per_seg          1D-array with shape (num_segments), indicates number of frames in each segment
    /n_frames                 number of frames in original video
    /picks                    positions of subsampled frames in original video
    /n_steps                  number of subsampled frames
    /gtsummary                1D-array with shape (n_steps), ground truth summary provided by user (used for training, e.g. maximum likelihood)
    /video_name (optional)    original video name, only available for SumMe dataset

Original videos and annotations for each dataset are also available in the authors' project webpages:

TVSum dataset: https://github.com/yalesong/tvsum

SumMe dataset: https://gyglim.github.io/me/vsum/index.html#benchmark

CSNet

We used the implementation of SUM-GAN method as a starting point to implement CSNet.

How to train

The implementation of CSNet is located under the directory csnet. Run main.py file with the configurations specified in configs.py to train the model.

SUM-Ind

Make splits

python create_split.py -d datasets/eccv16_dataset_summe_google_pool5.h5 --save-dir datasets --save-name summe_splits  --num-splits 5

As a result, the dataset is randomly split for 5 times, which are saved as json file.

Train and test codes are written in main.py. To see the detailed arguments, please do python main.py -h.

How to train

python main.py -d datasets/eccv16_dataset_summe_google_pool5.h5 -s datasets/summe_splits.json -m summe --gpu 0 --save-dir log/summe-split0 --split-id 0 --verbose

How to test

python main.py -d datasets/eccv16_dataset_summe_google_pool5.h5 -s datasets/summe_splits.json -m summe --gpu 0 --save-dir log/summe-split0 --split-id 0 --evaluate --resume path_to_your_model.pth.tar --verbose --save-results

Citations

@article{zhou2017reinforcevsumm, 
   title={Deep Reinforcement Learning for Unsupervised Video Summarization with Diversity-Representativeness Reward},
   author={Zhou, Kaiyang and Qiao, Yu and Xiang, Tao}, 
   journal={arXiv:1801.00054}, 
   year={2017} 
}

@inproceedings{DBLP:conf/aaai/JungCKWK19,
  author    = {Yunjae Jung and
               Donghyeon Cho and
               Dahun Kim and
               Sanghyun Woo and
               In So Kweon},
  title     = {Discriminative Feature Learning for Unsupervised Video Summarization},
  booktitle = {The Thirty-Third {AAAI} Conference on Artificial Intelligence, {AAAI}
               2019, The Thirty-First Innovative Applications of Artificial Intelligence
               Conference, {IAAI} 2019, The Ninth {AAAI} Symposium on Educational
               Advances in Artificial Intelligence, {EAAI} 2019, Honolulu, Hawaii,
               USA, January 27 - February 1, 2019},
  pages     = {8537--8544},
  publisher = {{AAAI} Press},
  year      = {2019},
  url       = {https://doi.org/10.1609/aaai.v33i01.33018537},
  doi       = {10.1609/aaai.v33i01.33018537},
  timestamp = {Wed, 25 Sep 2019 11:05:09 +0200},
  biburl    = {https://dblp.org/rec/conf/aaai/JungCKWK19.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}