LAUSR

Ellipse shapes taken by a projection lens encode rich semantic information, and ellipse detection is an important and convenient means to obtain spatial semantics from images. However, accurate and robust detectors for real-world images remain challenging. In this article, a shape-biased ellipse detection network with an auxiliary task is proposed to fulfill this requirement, which can learn and characterize low-level features more effectively. Our multitask-based architecture simultaneously performs elliptical bounding box detection, edge prediction, and parameter regression. First, since ellipses are represented as a set of arcs, we enhance the contact between low-level features and high-level features to better retain structural information. Second, ellipse edge prediction as an auxiliary task drives the backbone biased to ellipse shapes rather than others and makes more efficient use of limited datasets. In addition, by modeling the ellipse as a 2-D Gaussian distribution, Wasserstein distance is used as the loss function to minimize the distance between two Gaussian distributions. Experiments on real-world datasets show that the proposed EllipseNet achieves better performance than state-of-the-art methods. Extensive ablation studies validate the effectiveness of the proposed backbone module and auxiliary branch.