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

TL;DR

Edge detection suffers from costly manual labeling and annotation inconsistency. The paper introduces a self-supervised pipeline that transfers synthetic-edge labels to real images via Homography Adaptation and a dual-head model, SuperEdge, that predicts pixel-level and object-level edges using a three-stage training scheme: synthetic pretraining, real-scene pseudo-label generation with F(I,f) = (1/N_h) sum_{i=1}^{N_h} H_i^{-1} f(H_i(I)) and x = F(I,f) + G(I), and subsequent real-data training; a post-processing fusion O_fusion refines predictions. The method demonstrates strong cross-dataset performance, e.g., surpassing STEdge on BIPEDv2 and achieving notable gains by leveraging both pixel- and object-level cues, with losses L_pix and L_obj guiding learning. Overall, the approach reduces reliance on manual labels while achieving robust edge detection across diverse scenes, highlighting the potential of self-supervised, synthetic-to-real transfer for structured vision tasks.

Abstract

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

Figures (6)

• Figure 1: The edge-maps predictions from the proposed model in images acquired from internet.
• Figure 2: Training pipeline overview.Left: Our detector is first trained on an edge synthetic dataset of known ground truth. middle: Object-level and pixel-level pseudo ground truth of edge are generated on real images through methods such as L0 smoothing, canny, and homography adaptation. Right: Finally, the SuperEdge is trained on real images using the pseudo ground truth.
• Figure 3: Results of each detection head of SuperEdge on BIPED.
• Figure 4: P/R curves on BIPED dataset.
• Figure 5: P/R curves on different datasets. (a) BIPEDv2. (b) BSDS-RIND. (c) NYUD. (d) BSDS500.