Holistically-Nested Edge Detection

Holistically-Nested Edge Detection (HED) addresses edge detection through holistic image training and prediction combined with multi-scale, multi-level feature learning. The method performs image-to-image prediction using a deep learning model that leverages fully convolutional neural networks and deeply-supervised nets. Deep supervision on side responses guides the network to automatically learn rich hierarchical representations that help resolve the challenging ambiguity 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 processing images at 0.4 seconds each, orders of magnitude faster than earlier CNN-based edge detectors. It also produced encouraging results on additional boundary detection benchmarks such as Multicue and PASCAL-Context. This combination of accuracy and speed made holistic, end-to-end edge detection practical for broader computer vision use.