SketchVision: Bandwidth-Efficient Edge-Cloud Video Object Detection via Compressed Sketches
IEEE Transactions on Broadcasting · 2026-06 · 72 (2) · 545-557
DOI: 10.1109/tbc.2026.3680007 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Broadcasting, 2026-06. Verfasst von Hanling Wang und 6 weiteren Autor:innen. Die Arbeit nennt 73 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Basic and Frontier Research Project of the Peng Cheng Laboratory China, Project of the Department of Strategic and Advanced Interdisciplinary Research of Peng Cheng Laboratory, National Key Research and Development Program of China.
Abstract (Originalsprache)
A recent approach to video object detection captures video at the edge device and sends partial data to the cloud for inference, aiming to maximize accuracy within a latency bound. This approach addresses the limited computing power at the edge and the restricted bandwidth between edge and cloud. Existing solutions employ perceptual pruning, removing visually imperceptible details to reduce data transmission from edge to cloud, but this leads to suboptimal bandwidth savings. In this paper, we introduce SketchVision, a novel approach that leverages compressed sketches for semantic pruning to achieve bandwidth-efficient video object detection. SketchVision employs a lightweight region-aware sketch generator to extract key semantic line segments (which we refer to as sketches) from images. These extracted sketches then undergo optimal compression and encoding using a specially designed codec. Additionally, a context-aware object detection model is developed to leverage video-level conditional information, with a continuous learning mechanism to further enhance accuracy. Experiments on real-world datasets demonstrate that SketchVision achieves remarkable bandwidth savings of up to 52.7%, with a marginal 1.0% drop in accuracy compared to existing approaches.
Förderung
- Basic and Frontier Research Project of the Peng Cheng Laboratory China
- Project of the Department of Strategic and Advanced Interdisciplinary Research of Peng Cheng Laboratory
- National Key Research and Development Program of China
- Shenzhen Key Laboratory of Software Defined Networking
- Gabriel-Miro Muntean was supported by Research Ireland
Zitieren
APA
Wang, H., Li, Q., Jiang, Z., Bao, Z., Jiang, Y., Yuan, Z., & Muntean, G. (2026). SketchVision: Bandwidth-Efficient Edge-Cloud Video Object Detection via Compressed Sketches. IEEE Transactions on Broadcasting, 72(2), 545-557. https://doi.org/10.1109/tbc.2026.3680007
Vancouver
Wang H, Li Q, Jiang Z, Bao Z, Jiang Y, Yuan Z, et al. SketchVision: Bandwidth-Efficient Edge-Cloud Video Object Detection via Compressed Sketches. IEEE Transactions on Broadcasting. 2026;72(2):545-557. doi:10.1109/tbc.2026.3680007
BibTeX
@article{Wang2026,
author = {Wang, Hanling and Li, Qing and Jiang, Zhangzhang and Bao, Zepeng and Jiang, Yong and Yuan, Zhenhui and Muntean, Gabriel-Miro},
title = {SketchVision: Bandwidth-Efficient Edge-Cloud Video Object Detection via Compressed Sketches},
journal = {IEEE Transactions on Broadcasting},
year = {2026},
volume = {72},
number = {2},
pages = {545--557},
doi = {10.1109/tbc.2026.3680007}
}
EUVIMED