Model training and depth estimation method, device, storage medium and electronic device

Abstract

This specification discloses a model training and depth estimation method, apparatus, storage medium, and electronic device. At least two pairs of reference points are selected for each pixel in a sample image, wherein the reference point in each reference point pair is located on the same straight line as the pixel point. The pixel point and its reference point are then mapped into a three-dimensional space based on a depth map output by a depth estimation model to obtain three-dimensional coordinates, thereby obtaining at least two reference vectors with the reference points in the reference point pair as starting points and ending points. A predicted normal vector for the pixel point is determined based on the reference vectors, and a loss is constructed based on the difference between the predicted normal vector and the labeled normal vector corresponding to the pixel point for training, thereby improving the accuracy of the depth map output by the depth estimation model.

Images