Hi!
I am really interested in your work but I have little time to read the code. I have two questions about the evaluation Section in the paper.
1) How to select M=3 source views in one training pass? Randomly picked?
2) What's the value of M and how to select it in test/validation phase? Just send all of the source views into the network?
Ad 1) Yes, they are randomly sampled, but with with probability based on the overlap to the target image.
Ad 2) In the evaluation I use (mainly for memory reasons) max 5 views per 3D point.
Hi! I am really interested in your work but I have little time to read the code. I have two questions about the evaluation Section in the paper. 1) How to select M=3 source views in one training pass? Randomly picked? 2) What's the value of M and how to select it in test/validation phase? Just send all of the source views into the network?