Unmixing Convolutional Features for Crisp Edge Detection
This work provides an incremental improvement in localization accuracy for deep edge detectors, benefiting computer vision tasks that rely on precise edge detection.
This paper addresses the localization ambiguity in deep edge detectors, attributing it to feature mixing and side mixing in convolutional neural networks. The authors propose a Context-Aware Tracing Strategy (CATS) which includes a tracing loss for feature unmixing and a context-aware fusion block for side mixing. When integrated with a VGG16 backbone, CATS improved the F-measure (ODS) of RCF and BDCN deep edge detectors by 12% and 6% respectively on the BSDS500 dataset.
This paper presents a context-aware tracing strategy (CATS) for crisp edge detection with deep edge detectors, based on an observation that the localization ambiguity of deep edge detectors is mainly caused by the mixing phenomenon of convolutional neural networks: feature mixing in edge classification and side mixing during fusing side predictions. The CATS consists of two modules: a novel tracing loss that performs feature unmixing by tracing boundaries for better side edge learning, and a context-aware fusion block that tackles the side mixing by aggregating the complementary merits of learned side edges. Experiments demonstrate that the proposed CATS can be integrated into modern deep edge detectors to improve localization accuracy. With the vanilla VGG16 backbone, in terms of BSDS500 dataset, our CATS improves the F-measure (ODS) of the RCF and BDCN deep edge detectors by 12% and 6% respectively when evaluating without using the morphological non-maximal suppression scheme for edge detection.