Fig. 2: Stereo depth estimation results.
From: A multichannel optical computing architecture for advanced machine vision
a Network architecture of the Monet for stereo depth estimation, composed of an interference pattern generation module followed by a depth regression module. The former module encodes the stereo correspondence into the interference patterns, and the latter module adopts a U-net-like structure to estimate the depth information from the patterns. b Projection function for stereo depth estimation. The spatial position of the left-view image is fixed and the right-view image is shifted to the right according to the depth range. c Interference patterns with 5 shift distances (presented below each interference pattern). d A large area from the test set. Stereo depth estimation results of 4 representative blocks (labelled using yellow boxes) are shown on the right. e Stitched depth map of the large area in subfig. d The depth value is converted to the height relative to the ground for a more intuitive understanding. The heights along two lines L1 and L2 are presented in the middle. The depth-error distribution of the whole area is shown on the right. f Height-coloured 3d model of the large area. Norm normalized, GT ground-truth, px pixel
