A Novel Ellipse Object Detection Network Based on Adaptive Point Set

This is the authors' abstract. We don't add key points for this paper.

Object detection is a crucial task in the field of computer vision, and circular objects are commonly encountered in various industrial and transportation domains. However, in real-world scenarios, captured objects often appear as ellipses due to the influence of shooting angles. Hence, ellipse object detection has broader applicability compared to circular object detection. In recent years, ellipse detection algorithms have mainly focused on traditional methods, and applying deep learning techniques to ellipse object detection remains challenging. In this regard, we propose a novel ellipse detection network that combines a verification branch with a regression branch, along with an improved feature point evaluation method. This allows the rotation-based object detector to better learn and represent the characteristics of ellipses. Compared to existing algorithms, our trained model demonstrates superior performance on limited datasets, excelling in multiple evaluation metrics.