A method and apparatus for estimating the depth of a binocular image, a device, a program, and a medium are provided. The method includes: inputting a to-be-processed binocular image into a predetermined neural network to obtain a parallax image and a confidence score image thereof, where a discrete parameter of each pixel in a preset distribution in the parallax image has a negative correlation with a confidence score of the pixel; and obtaining a depth image of the to-be-processed binocular image according to the parallax image.