One-Stage Wireframe Parsing in Fish-Eye Images

This paper presents a simple yet efficient algorithm for detecting wireframes in a fish-eye image. Given a fish-eye image, our objective is to extract the semantically and geometrically salient lines and their corresponding junctions directly and without any undistortion processing. To this end, we introduce the circular-arc-based representation, converting the distorted line segments into a part of a circle by drawing inspiration from the projection characteristics of fish-eye lenses. Then a compact and end-to-end trainable model that tailors the redundant multi-module is realized and is used to directly output the vectorized wireframe in a one-stage fashion. Compared with previously considered the state-of-art two-stage method for line segment detection in the fish-eye image, our method achieves an average improvement of 2.8 in structural average precision, and boasts a remarkable increase in inference speed. Moreover, our method demonstrates universality when handling images captured by standard cameras, outperforming existing straight wireframe parsing methods.