Iterative Residual Network for Structured Edge Detection

Edge detection aims to find visually distinctive edges or boundaries in input images. Edge detection has made significant progress with the help of deep Convolutional Networks (ConvNet). Most ConvNet-based edge detectors predict each pixel independently and ignore the inherent correlations between pixels. However, structured cues in input images are critical to learn a good edge detector. To this end, we propose a novel Iterative Residual Holistically-nested Edge Detection (IRHED) network. IRHED incorporates multi-scale features from the hierarchy of the network, and learns to iteratively refine the output boundary map in a deeply supervised manner. In this way, global structural cues, such as object shape, are learned implicitly, thus edges can be effectively distinguished. Extensive experiments demonstrate that IRHED achieves state-of-the-art results on the widely used BSDS500 dataset. We also show the benefit of structured edge map for higher-level task, such as object proposal generation.