Video Tiny-Object Detection Guided by the Spatial-Temporal Motion Information
Abstract
Detecting tiny/small objects (e.g., drone targets) in videos is highly desired in many realistic scenarios. Nevertheless, current object detection algorithms can hardly recognize tiny targets against extremely complex backgrounds. To address this problem, we propose a motion-guided video tiny-object detection method (MG-VTOD), in which the spatial-temporal motion strength maps play an important role in object searching and locating. Inspired by the biological retinal structure, we compute the motion strength using a sequential frame cube that has been aligned and registered. Subsequently, the motion strength maps are employed to enhance the potential areas of the moving targets, thereby facilitating the target detection procedure. Experimental results obtained on the Anti-UAV-2021 dataset validate that the proposed MG-VTOD method significantly outperforms the competing object detection methods.
Cite
Text
Yang et al. "Video Tiny-Object Detection Guided by the Spatial-Temporal Motion Information." IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2023. doi:10.1109/CVPRW59228.2023.00307Markdown
[Yang et al. "Video Tiny-Object Detection Guided by the Spatial-Temporal Motion Information." IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2023.](https://mlanthology.org/cvprw/2023/yang2023cvprw-video/) doi:10.1109/CVPRW59228.2023.00307BibTeX
@inproceedings{yang2023cvprw-video,
title = {{Video Tiny-Object Detection Guided by the Spatial-Temporal Motion Information}},
author = {Yang, Xin and Wang, Gang and Hu, Weiming and Gao, Jin and Lin, Shubo and Li, Liang and Gao, Kai and Wang, Yizheng},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops},
year = {2023},
pages = {3054-3063},
doi = {10.1109/CVPRW59228.2023.00307},
url = {https://mlanthology.org/cvprw/2023/yang2023cvprw-video/}
}