DSD: Depth Structural Descriptor for Edge-Based Assistive Navigation

Abstract

Structural edge detection is the task of finding edges between significant surfaces in a scene. This can underpin many computer vision tasks such as sketch recognition and 3D scene understanding, and is important for conveying scene structure for navigation with assistive vision. Identifying structural edges from a depth image can be challenging because surface structure that differentiates edges is not well represented in this format. We derive a depth input encoding, the Depth Surface Descriptor (DSD), that captures the first order properties of surfaces, allowing for improved classification of surface geometry that corresponds to structural edges. We apply the DSD feature to salient edge detection on RGB-D images using a fully convolutional neural network with deep supervision. We evaluate our method on both a new RGB-D dataset containing prosthetic vision scenarios, and the SUNRGBD dataset, and show that our approach produces improved performance compared to existing methods by 4%.