# Supervised Video Summarization via Multiple Feature Sets with Parallel Attention

**Authors:** Junaid Ahmed Ghauri, Sherzod Hakimov, Ralph Ewerth  
**arXiv:** [2104.11530](https://arxiv.org/abs/2104.11530)  
**Date:** 2021-04-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2104.11530v2/

## Key points

- The paper addresses the task of assigning importance scores to frames or short segments in a video for summarization.
- Existing methods rely on a single source of visual features, limiting their effectiveness.
- The authors propose a novel model architecture that combines three feature sets representing visual content and motion.
- The proposed architecture incorporates an attention mechanism to capture relevant information and improve prediction of importance scores.
- Comprehensive experimental evaluations are conducted on SumMe and TVSum benchmark datasets.
- Methodological issues with previous work using these datasets are identified, and a fair evaluation scheme is presented for future research.
- Results show significant improvements over state-of-the-art methods for SumMe dataset, and comparable performance for TVSum dataset.
- The paper contributes to advancing the field by addressing methodological issues and providing a fair evaluation scheme.

## Summary

The paper titled "Supervised Video Summarization via Multiple Feature Sets with Parallel Attention" addresses the challenging task of assigning importance scores to frames or short segments in a video for the purpose of summarization. The existing methods in this field rely on a single source of visual features, which may limit their effectiveness. To overcome this limitation, the authors propose a novel model architecture that combines three feature sets representing visual content and motion. The proposed architecture incorporates an attention mechanism before fusing the motion features and features derived from an image classification model, which represent the static visual content. This attention mechanism helps in capturing relevant information and improving the prediction of importance scores. To evaluate the performance of their approach, comprehensive experimental evaluations are conducted on two well-known benchmark datasets: SumMe and TVSum. In doing so, the authors also identify methodological issues with how previous work has used these datasets and present a fair evaluation scheme with appropriate data splits that can be utilized in future research. The results obtained from using static and motion features with parallel attention mechanism show significant improvements over state-of-the-art methods for SumMe dataset. For TVSum dataset, the proposed approach achieves comparable performance to the current state-of-the-art methods. In conclusion, this paper presents a novel model architecture for supervised video summarization that combines multiple feature sets and utilizes an attention mechanism. The experimental evaluations demonstrate its effectiveness on two benchmark datasets. The findings contribute to advancing the field by addressing methodological issues and providing a fair evaluation scheme for future research in video summarization.