Figure 5. The architecture of the proposed DDDNet. With the input DP image pair, the DepthNet first estimates a coarse inverse depth map
Dˆ . Then, we combine the coarse and blurred inverse depth map with B , and feed them to the DeblurNet. The deblurred image and
c {L,R}
estimated inverse depth map are fed to our DP simulator to synthesize a DP image pair. The synthesized DP images are compared to the
inputs using our reblur loss (Eq. (11)), to regularise the DDDNet in training. Note that, the ground truth inverse depth maps and sharp
images are used as supervision signals in training. (Best viewed in colour on screen.)