Learning in Low-level Vision

One of the key findings of CNNs was that learned visual features are superior to hand-crafted features. However, the vast majority of works start from 8-bit RGB data at fixed resolution as raw image data without further reflecting on the pre-processing that is already involved to prepare data this way. Obviously, one could go much further down to the physical measurement layer to extract data, but will this continue to improve performance? Will the amount of required data increase? Can we still maintain hardware indepdendence if going closer to the physical measurement layer? During the talk, some results related to those questions will be discussed and reflected, but most importantly, the modeling of several steps between the physical measurement layer and standardized image data will be discussed.