This repository contains the official implementation for the paper Breaking The Synopsis Trade-off: Preserving Event Cohesion Via Community-aware Graph Packing. Published in the International Journal of Web Research (IJWR), 2026.
🔗 Paper: https://ijwr.usc.ac.ir/article_244594.html Our work introduces a novel framework for video synopsis that significantly improves both summarization efficiency and the narrative coherence of the final video.
Unlike traditional methods that often create chaotic summaries by arranging objects based solely on individual importance, our approach models the video content as a heterogeneous graph. By identifying and preserving "communities" of interacting objects, we generate a synopsis that is not only compact but also logically and visually coherent.
Our pipeline consists of several key stages:
- Dynamic Preprocessing: Includes adaptive background generation, robust tracklet merging using visual features (ResNet-50), and intelligent data cleaning.
- Heterogeneous Graph Modeling: Objects and their complex spatio-temporal and visual interactions are modeled using a rich heterogeneous graph.
- GNN-based Importance Scoring: A Graph Attention Network (GAT) is trained to learn context-aware importance scores for each object.
- Community-Aware Packing: A novel packing algorithm that uses the detected communities (events) and their GNN scores to schedule objects, preserving the narrative flow.
- Advanced Evaluation: We introduce a new metric, the Community Cohesion Ratio (CCR), to quantitatively measure the narrative coherence of a synopsis.
Our method was evaluated on the SynoClip dataset and compared againstthe existing benchmark. The results demonstrate a significant performance leap, simultaneously improving both the Frame Rate (Fr) and the Collision Area (CA), effectively breaking the traditional trade-off between compression and visual quality.
Comparison of our method (blue) against the existing benchmark (red), averaged over five videos. The ideal point is the bottom-left corner.
The experiments in this paper were conducted on the SynoClip dataset, a public benchmark introduced by Malekpour et al. in their 2024 paper, "A Low-Computational Video Synopsis Framework with a Standard Dataset".
You can download the full SynoClip dataset from the original source linked below:
- Download SynoClip Dataset: [SynoClip Dataset]
To run this project, first clone the repository and install the required dependencies.
git clone https://github.com/Hosein541/Community-Aware-Video-Synopsis.git
cd Community-Aware-Video-Synopsis
pip install -r requirements.txtThe entire pipeline is contained within the main Jupyter Notebook.
- Download the dataset and place the videos in your working directory.
- Open the
community_aware_video_synopsis.ipynbnotebook. - In the first code cell, set the
video_pathto your desired input video and choose adensification_factor. - Run all cells in the notebook in order. The final video synopsis and all evaluation results will be generated and saved in the project folder.
If you find this work useful in your research, please consider citing our paper:
@article {
author = {Mohamadi, Hosein and Shayegan, Mohammad Javad},
title = {Breaking The Synopsis Trade-off: Preserving Event Cohesion Via Community-aware Graph Packing},
journal = {International Journal of Web Research},
volume = {9},
number = {2},
pages = {91-107},
year = {2026},
publisher = {University of Science and Culture},
issn = {2645-4335},
eissn = {2645-4343},
doi = {10.22133/ijwr.2026.564052.1322},
abstract = {Video synopsis condenses long surveillance recordings into concise, informative summaries. Existing methods often sacrifice narrative continuity, dispersing related events across the timeline to achieve higher compression. In this work, we reframe the task by modeling video content as a heterogeneous interaction graph. A Graph Neural Network (GNN) computes context-aware refinements of heuristic importance scores for objects, capturing the nuanced relationships that define complex scenes. At the core of our framework lies a community-aware packing algorithm, which schedules event-based communities rather than individual objects, thereby preserving temporal and semantic integrity. Extensive experiments show that our approach surpasses the conventional compression-quality trade-off, achieving 58% higher compression rates than the existing benchmark method while reducing visual clutter. We introduce the Community Cohesion Ratio (CCR), a task-specific evaluation metric designed to quantify the preservation of relative temporal relationships among objects belonging to the same event community. Using this complementary metric, our approach achieves an 88% lower CCR value than the standard packing method, indicating improved preservation of temporal community structure.},
keywords = {Video Synopsis,Event Cohesion,Graph Packing,community detection,graph neural networks},
url = {https://ijwr.usc.ac.ir/article_244594.html},
eprint = {https://ijwr.usc.ac.ir/article_244594_f494fa5ddcfb713c3f9f35566ad0f080.pdf}
}