sxyu
Reproducing two object scene evaluation results
I was wondering if you could please provide some more information concerning the evaluation protocol for the two object scene evaluation results. I see that there's a file named viewlist/2obj_eval_views.txt with some indices in it, but it doesn't appear to be in the format used for the other evaluations.
For these tests, are you using 1-3 specific input view indices to render a specific target viewpoint? Or are you measuring the results when rendering all non-input views?
I also noticed that the paper says that there are 50 testing views, but the test data split appears to only have 20 views, as in the training data (Sec. B.2.4). Is the data that's provided the same as what's being used for these evaluations?
Same here, I tried the official ckpt, and test with view '10', the PSNR is around 16, which is significantly different from the 20.15 PSNR reported in the paper. Kind of confusing.
