EPINet+BpCNet
Take your model further: a general optimization network for light field depth estimation via bad pixel optimization
We propose a general optimization network for light field depth estimation by optimizing bad pixels-BpcNet, which can almost further improve the results of existing algorithms. Given an initial depth map (possibly from other algorithms), BpCNet assumes that the error of all bad pixels on it is within $[-\delta,\delta]$, then models it accordingly.
By searching within a limited range, BpCNet will optimize the bad pixels for smaller errors. We conduct experiments on multiple algorithms to verify the effectiveness of BpCNet and this is for EPINet.
full grid
python
Ubuntu16.04, 1080ti
