SuperEdge: Towards a Generalization Model for Self-Supervised Edge Detection

• DOI: 10.48550/arxiv.2401.02313
• Publication Date: 2024-01-01
• Authors: Leng Kai; Zhijie Zhang; Jie Liu; Zeyd Boukhers; Wei Sui; Cong Yang; Zhijun Li
• Keywords: Leverage (statistics); Granularity

Edge detection is a fundamental technique in various computer vision tasks. Edges are indeed effectively delineated by pixel discontinuity and can offer reliable structural information even textureless areas. State-of-the-art heavily relies on pixel-wise annotations, which labor-intensive subject to inconsistencies when acquired manually. In this work, we propose novel self-supervised approach for edge that employs multi-level, multi-homography transfer annotations from synthetic real-world datasets. To fully leverage the generated developed SuperEdge, streamlined yet efficient model capable of concurrently extracting edges at pixel-level object-level granularity. Thanks training, our method eliminates dependency manual annotated labels, thereby enhancing its generalizability across diverse Comparative evaluations reveal SuperEdge advances detection, demonstrating improvements 4.9% ODS 3.3% OIS over existing STEdge BIPEDv2.