Robust Use of Semantic Segmentation for Depth and Disparity Estimation

摘要

This disclosure relates to techniques for generating robust depth estimations for captured images using semantic segmentation. Semantic segmentation may be defined as a process of creating a mask over an image, wherein pixels are segmented into a predefined set of semantic classes. Such segmentations may be binary (e.g., a ‘person pixel’ or a ‘non-person pixel’) or multi-class (e.g., a pixel may be labelled as: ‘person,’ ‘dog,’ ‘cat,’ etc.). As semantic segmentation techniques grow in accuracy and adoption, it is becoming increasingly important to develop methods of utilizing such segmentations and developing flexible techniques for integrating segmentation information into existing computer vision applications, such as depth and/or disparity estimation, to yield improved results in a wide range of image capture scenarios. In some embodiments, an optimization framework may be employed to optimize a camera device's initial scene depth/disparity estimates that employs both semantic segmentation and color regularization in a robust fashion.