taskgrouping
taskgrouping copied to clipboard
tstandley
Question about data pre-processing.
Hi, in the taskonomy_loader.py, I see different pre-processing operations are performed according to different tasks, i.e., from Line 171 to Line 208. Can you explain a bit about the underlying principle for these operations? Or is there any reference ? Thank you so much!
Much of this is to normalize the outputs. Each task has a different range of outputs and I generally standardize them to have a mean of ~zero and a standard devidation of ~1.
I also flip images as part of my data augmentation. This is easy to do for most tasks, but for normals (for example) you have to reverse the direction of the normal whenever you filp the image otherwise you are learning garbage.
There may be some other stuff I did. It's been quite some time since I wrote that code, but if you have a question about a specific loading step, I might be able to remember.
@tstandley Thanks for the reply, it's very helpful! I noticed that in Line 173, a mak is generated, where the original values less than 2^13 are set to be valid. However, the maximum value of the depth_zbuffer is 2^16-1, why the values between 2^13 and 2^16-1 are invalid?
Not sure. My guess is that was the maximum valid entry I observed in the data.