A new baseline for edge detection: Make encoder–decoder great again

Abstract

The performance of deep learning based edge detectors has surpassed human performance, but the huge computational cost and complex training strategies hinder their further development and application. In this paper, we alleviate these complexities with a vanilla encoder–decoder based detector. Firstly, we design a bilateral encoder to decouple the extraction process of spatial features and semantic features. As the spatial branch no longer guides the semantic branch, feature richness can be reduced, enabling a more compact model design. We propose a cascaded feature fusion decoder, where the spatial features are progressively refined by semantic features. The refined spatial features are the only basis for generating the edge map. The coarse original spatial features and semantic features are avoided from direct contact with the final result. So the noise in the spatial features and the location error in the semantic features can be suppressed in the generated edge map. The proposed New Baseline for Edge Detection (NBED) achieves superior performance consistently across multiple edge detection benchmarks, even compared with those methods with huge computational costs and complex training strategies. The ODS of NBED on BSDS500 is 0.838, achieving state-of-the-art performance. Our study highlights that high-quality features are key to modern edge detection, and encoder–decoder based detectors can achieve excellent performance without complex training or heavy computation. Furthermore, we take retinal vessel segmentation as an example to explore the application of NBED in downstream tasks. The code is available at https://github.com/Li-yachuan/NBED.