PGL-SUM pretrained models

This dataset contains pretrained models of the PGL-SUM network architecture for video summarization, that is presented in our work titled “Combining Global and Local Attention with Positional Encoding for Video Summarization”, in Proc. IEEE ISM 2021. 

This work introduces a new method for supervised video summarization, which aims to overcome drawbacks of existing RNN-based summarization architectures that relate to the modeling of long-range frames' dependencies and the ability to parallelize the training process. The proposed PGL-SUM network architecture relies on the use of self-attention mechanisms to estimate the importance of video frames. Contrary to previous attention-based summarization approaches that model the frames' dependencies by observing the entire frame sequence, our method combines global and local multi-head attention mechanisms to discover different modelings of the frames' dependencies at different levels of granularity. Moreover, the utilized attention mechanisms integrate a component that encodes the temporal position of video frames - this is of major importance when producing a video summary. Experiments on two benchmarking datasets (SumMe and TVSum) demonstrate the effectiveness of the proposed model compared to existing attention-based methods, and its competitiveness against other state-of-the-art supervised summarization approaches.

File format

The provided “pretrained_models.zip“ file contains two sets of pretrained models of the PGL-SUM network architecture. After downloading and unpacking this file, in the created “pretrained_models” folder you will find the following sub-directories:

• table3_models,

• table4_models

The sub-directory “table3_models“ contains models of the PGL-SUM network architecture that have been trained in a single-batch mode and were manually selected based on the observed summarization performance on the videos of the test set. The average performance of these models (over the five utilized data splits) is reported in Table III of [1].

The sub-directory “table4_models“ contains models of the PGL-SUM network architecture that have been trained in a full-batch mode and were automatically selected after the end of the training process based on the recorded training losses and the application of the designed model selection criterion (described in Section IV.B of our paper). The average performance of these models (over the five utilized data splits) is reported in Table IV of [1].

Each of these sub-directories contains the pretrained model (.pt file), for:

• Each utilized benchmarking dataset: {SumMe, TVSum}

• And each utilized data-split: {0, 1, 2, 3, 4}

The naming of each of the provided .pt files indicates the training epoch associated with the selected pretrained model.

Finally, the data-splits we used for performing inference on the provided pretrained models, and the source code that can be used for training your own models of the proposed PGL-SUM network architecture, can be found at: https://github.com/e-apostolidis/PGL-SUM.

License and Citation

This dataset is provided for academic, non-commercial use only. If you find this dataset useful in your work, please cite the following publication where it is introduced:

[1] E. Apostolidis, G. Balaouras, V. Mezaris, I. Patras, "Combining Global and Local Attention with Positional Encoding for Video Summarization", Proc. 23rd IEEE Int. Symposium on Multimedia (ISM), Dec. 2021. Software available at: https://github.com/e-apostolidis/PGL-SUM

Acknowledgements

This work was supported by the EU Horizon 2020 programme under grant agreement H2020-832921 MIRROR, and by EPSRC under grant No. EP/R026424/1.