Closed xcharxlie closed 1 year ago
I use 640 as the sampling interval to save the evaluation time. I have also tried evaluating on the whole testset. This generally gives similar or even slightly better performance despite much longer evaluation time.
And most of the results listed in the big table are probabilistic right, unless the ones are stated explicitly.
In my experience, there is very little variance in the results of each evaluation. The results reported in the paper can be reproduced using the scripts and checkpoints provided in this repo.
Thank you so much! I'm also looking forward to the 3dhp part of code for result comparison.
Did you use 640 as the sampling interval in evaluation or you evaluated on the whole test dataset?