View in EDS

Dynamic video summarization using handcrafted features to complement publicly available datasets.

Saved in:
Bibliographic Details
Title: Dynamic video summarization using handcrafted features to complement publicly available datasets.
Authors: Solatidehkordi, Zahra1 (AUTHOR), Shanableh, Tamer1 (AUTHOR) tshanableh@aus.edu
Source: Multimedia Tools & Applications. Nov2025, Vol. 84 Issue 37, p46457-46478. 22p.
Subjects: Video summarization, Motion estimation (Signal processing), Data integration
Abstract: In video summarization, four datasets (TvSum, SumMe, OVP, and YouTube) are typically used for training and testing. In this study, we supplement these datasets with novel features based on High Efficiency Video Codec (HEVC) video coding and motion estimation and compensation. Although HEVC coding variables offer valuable information, they are frequently overlooked in deep learning solutions for video analysis. Thus, we introduce a low-level HEVC feature set suitable for dynamic video summarization. Additionally, we supplement the datasets with additional feature vectors based on CNN embeddings of Window-based Accumulated Image Differences with motion estimation and compensation (WAID-MC). We integrate our two proposed feature sets using feature vector fusion and importance score fusion. Furthermore, we enhance the OVP and YouTube datasets by adding ground truth keyshots, importance scores, and user summaries. In the experimental results section, we compare our proposed solutions with four existing approaches, integrating our features and fusion techniques into their codebases. Our findings indicate that in a minimum of three out of the four datasets, the F1 scores achieved by our proposed methodologies are superior to those of existing studies. Moreover, the experimental results predominantly highlight feature vector fusion as superior to importance score fusion. In many instances, the fusion of the WAID-MC features with the HEVC features yields the best F1 scores. In some cases, the utilization of HEVC features alone results in higher F1 scores compared to existing approaches. [ABSTRACT FROM AUTHOR]
Copyright of Multimedia Tools & Applications is the property of Springer Nature and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:In video summarization, four datasets (TvSum, SumMe, OVP, and YouTube) are typically used for training and testing. In this study, we supplement these datasets with novel features based on High Efficiency Video Codec (HEVC) video coding and motion estimation and compensation. Although HEVC coding variables offer valuable information, they are frequently overlooked in deep learning solutions for video analysis. Thus, we introduce a low-level HEVC feature set suitable for dynamic video summarization. Additionally, we supplement the datasets with additional feature vectors based on CNN embeddings of Window-based Accumulated Image Differences with motion estimation and compensation (WAID-MC). We integrate our two proposed feature sets using feature vector fusion and importance score fusion. Furthermore, we enhance the OVP and YouTube datasets by adding ground truth keyshots, importance scores, and user summaries. In the experimental results section, we compare our proposed solutions with four existing approaches, integrating our features and fusion techniques into their codebases. Our findings indicate that in a minimum of three out of the four datasets, the F1 scores achieved by our proposed methodologies are superior to those of existing studies. Moreover, the experimental results predominantly highlight feature vector fusion as superior to importance score fusion. In many instances, the fusion of the WAID-MC features with the HEVC features yields the best F1 scores. In some cases, the utilization of HEVC features alone results in higher F1 scores compared to existing approaches. [ABSTRACT FROM AUTHOR]
ISSN:13807501
DOI:10.1007/s11042-025-20988-8