Point-to-RBox Network for Oriented Object Detection via Single Point Supervision
Yucheng Wang (WuHan University),* Chu He (Wuhan University), Xi Chen (Wuhan university)
The 34th British Machine Vision Conference
Abstract
The Rotated Bounding Boxes used in Oriented object detection are labor-intensive and time-consuming to annotate manually. Unlike rotated boxes with fine granularity, point-level annotations only provide a single point for each object as supervision, greatly reducing the annotation burden. In this paper, we formalize the problem as using point annotations to generate high-quality pseudo rotated boxes that can be used to train existing detectors. To address the core challenge of generating pseudo rotated boxes, we propose the Point-to-RBox (P2RBox) network. First, we introduce a coarse-to-fine strategy to generate precise pseudo rotated boxes. Second, to account for objects with arbitrary orientation, we design a three-stream detection head guided by orientation-sensitive features in P2RBox to select the best pseudo rotated box. The extensive experiments on the DOTA and DIOR-R datasets indicate that the pseudo rotated boxes generated by P2RBox are viable substitutes for manually annotated rotated boxes. Using pseudo rotated boxes, a fully-supervised object detector can attain more than 90\% of the performance achieved by the same detector trained with manually annotations. In addition, our method not only outperforms image-level weakly supervised detectors but also exhibits competitive performance compared to the fully supervised detectors.Video
Citation
@inproceedings{Wang_2023_BMVC,
author = {Yucheng Wang and Chu He and Xi Chen},
title = {Point-to-RBox Network for Oriented Object Detection via Single Point Supervision},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0323.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by