Supervised data driving-based monocular video depth estimating method

Abstract

The invention discloses a supervised data driving-based monocular video depth estimating method comprising the following steps: (1) a sample video sequence and a corresponding depth sequence are obtained as a training data set; (2) a tracking-based superpixel segmentation method is used for segmenting the training data set, and characteristics of each segmented unit are extracted; (3) a network model combining a convolution nerve network and a space-time condition random field; (4) the training data set, the segmentation result and the corresponding characteristics are used for training a depth space-time convolution nerve network field model; (5) a video sequence to be estimated is segmented, and characteristics of each segmented unit are extracted; (6) the video sequence to be estimated, the segmentation result and the corresponding characteristics are input into the trained model, and a depth sequence can be obtained. The supervised data driving-based monocular video depth estimating method is based on consideration on space-time consistency and hierarchical relation accuracy, and monocular stereoscopic video quality can be improved.

Images