Depth prediction from image data using statistical models

摘要

The present invention discloses a system and method for predicting depth from color image data using a statistical model such as a convolutional neural network (CNN). The model is trained on a binocular stereo image pair, allowing depth data to be predicted from a single source curler image. The model is trained for each image of the input binocular stereo pair to predict a corresponding disparity value that, when applied to the image, allows reconstruction of other images. The model is updated based on a cost function that increases the consistency between predicted parallax values for each image of a stereo pair.