Line2Rbox: Line-supervised Oriented Object Detection
Keywords: Line-supervised, Oriented Object Detection, Remote sensing images, Weakly-supervised
Abstract: Oriented object detection is crucial for complex scenes such as aerial images and industrial inspection, providing precise delineation by minimizing background interference. Recently, the weakly-supervised oriented object detection has gaining attention due to its cost-effectiveness. However, the majority of existing weakly-supervised methods are either point-supervised or HBox-supervised, which presents a challenge in achieving an optimal balance between annotation cost and detection performance. In response, we introduce a novel form of line annotation, which is intermediate between point-level and plane-level annotation. Based on this, we present L2RBox, an end-to-end anchor-free detector that is the first line-supervised method for oriented object detection. The fundamental objective of the L2RBox is to utilise line labels for the completion of label assignment and the calculation of loss. In particular, the line is mapped to the corresponding circle domain, which is then used to select training samples and calculate the center-ness target by the minimum circumscribed rectangle of the circle in the direction of the line. The regression loss that we propose is designed to support the line as an optimisation target. It comprises four components, namely scale loss $L_s$, height loss $L_h$, position loss $L_p$ and angle loss $L_a$.
Extensive experimentation on DOTA-v1.0 and DIOR-R has demonstrated that our L2RBox significantly outperforms point-supervised methods, while requiring only a slight increase in labeling costs. It is also noteworthy that the proposed approach also demonstrates a slight performance advantage over the fully-supervised FCOS in certain categories.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 128
Loading