Self-supervised training for depth estimation models using depth hints

摘要

PROBLEM TO BE SOLVED: To provide a training method of a depth estimation model for generating a depth map. SOLUTION: For each image pair, the depth prediction for the first image is determined by the depth estimation model, the depth hint is acquired, and the depth is projected in the first projection of the second image onto the first image. Generate a predictive composite frame, then generate a hinted composite frame based on depth hints on the second projection, calculate the primary loss using the composite frame, and calculate the hinted loss using the hinted composite frame. Calculate and calculate the total loss for the image pair based on the pixel-by-pixel determination. Here, if the hinted loss is less than the primary loss, the total loss includes the primary loss and the supervised depth loss between the depth prediction and the depth hint. The depth estimation model is trained by minimizing the total loss of the image pair. [Selection diagram] FIG. 4