Holistically-Nested Edge Detection

Holistically-Nested Edge Detection (HED) is an edge detection algorithm built to address two issues together: holistic image training and prediction, and multi-scale, multi-level feature learning. It performs image-to-image prediction using a deep learning model that combines fully convolutional neural networks with deeply-supervised nets, automatically learning rich hierarchical representations guided by deep supervision on side responses to resolve the ambiguity inherent in edge and object boundary detection.

HED significantly advanced the state of the art, reaching an ODS F-score of 0.790 on the BSDS500 dataset and 0.746 on the NYU Depth dataset, while running at 0.4 seconds per image, orders of magnitude faster than some earlier CNN-based edge detectors. It also showed encouraging results on other boundary detection benchmarks such as Multicue and PASCAL-Context, demonstrating the value of holistic prediction and deep supervision for boundary detection.