Skip to search boxSkip to navigationSkip to main content
Logo Gonzaga University

DSTG-VS: Dynamic spatio-temporal graph network with global-local feature fusion for video summarization

Research Output:
Contribution to journal
Article
Research
 Peer-review

Abstract

Video summarization aims to extract concise and informative representations of long videos, saving storage space and enabling efficient processing. However, conventional sequence-based models often fail to capture non-consecutive semantic correlations, while existing graph-based approaches are hindered by rigid, static topologies that cannot adapt to varying narrative rhythms. Furthermore, deep graph layers frequently suffer from the “over-smoothing” effect, which washes out fine-grained local motion details and causes an imbalance in feature integration. To directly address these limitations, we propose a novel dynamic spatio-temporal graph network with global-local feature fusion for video summarization (DSTG-VS), which transforms video frame features into dynamic graphs. Semantic correlations between frame pairs are dynamically determined by temporal decay and frame similarity matrices. A dual-path fusion mechanism is further designed: global features (from adaptive average pooling) mitigate over-smoothing in graph convolutions, while local features preserve critical fine-grained motion details through adaptive contribution balancing via a dynamic gating module. On SumMe and TVSum datasets, the proposed DSTG-VS achieves 64.5% and 61.7% F-score, respectively, outperforming the state-of-the-art method (Wang et al.) on SumMe by up to 7.1%, and achieving a similar F-Score to the state-of-the-art method TCVT on TVSum with only 20.7% of its parameters. DSTG-VS achieves superior efficiency and effectiveness on continuous unedited videos, establishing a new state-of-the-art for graph-based spatio-temporal video modeling.

Access to documents

Bibliographic Information

Output type

Research Output:
Contribution to journal
Article
Research
 Peer-review

Original language

American English

Journal (Volume, Issue Number)

Pattern Recognition

Publication milestones

  • Published - 10/2026

Publication status

Published - 10/2026

Publication IDs

  • Scopus: 105033231422