As Refik indicated, usability of different datasets in a merged form is investigated. As far as I saw from papers and implementations , different datasets are seperately used for training and tested with extracted models. However, for particularly make3D(outdoor) and NYU datasets(indoor), depth values are given in meters for both(these datasets are common benchmarks for depth estimation).I have also checked some images in matlab for both datasets and observed that the values of pixels confirms this fact. Overall, it is okay to shuffle them without any scaling(at least in theory). However, I am still skeptical to do so, since it is not a common application also they are different data come from different sensors. I thought it would be nice to discuss pros and cons to do so. Also, different input sizes is another issue.
Can you also check these datasets for being in the same page?
As Refik indicated, usability of different datasets in a merged form is investigated. As far as I saw from papers and implementations , different datasets are seperately used for training and tested with extracted models. However, for particularly make3D(outdoor) and NYU datasets(indoor), depth values are given in meters for both(these datasets are common benchmarks for depth estimation).I have also checked some images in matlab for both datasets and observed that the values of pixels confirms this fact. Overall, it is okay to shuffle them without any scaling(at least in theory). However, I am still skeptical to do so, since it is not a common application also they are different data come from different sensors. I thought it would be nice to discuss pros and cons to do so. Also, different input sizes is another issue.
Can you also check these datasets for being in the same page?