kcyt
commented
1 year ago We do realize that there is a number of varying factors, such as data augmentation usage, cleaning of THuman2.0 meshes, and the yaw angles used, that make it hard to obtain the exact scores.
We will provide an update on this. Concretely, we will provide steps to replicate the scores and update the scores because right now we no longer use 10 random yaw angles but a frontal angle that is then fanned out in both clockwise and counter-clockwise directions. We will also fix the other varying factors.
Also, our PIFu is taken from the original PIFu repository, and spatial sampling is used in training the vanilla PIFu.
Hi, I am trying to attain the CD and P2S scores shown in the paper for PIFu trained on THuman2.0 dataset for benchmarking purposes, however I have not been successful. I used the HGPIFuNetwNML model with spatial sampling, but I only managed to achieve ~4.0e-4 CD after 500 epochs, whereas PIFu was able to achieve 3.135e-4 CD in the paper.
May I have some information about the configuration used? 1) What are the 10 yaw angles used in the image rendering? 2) Is the vanilla PIFu used in the paper taken directly from the original PIFu repository, or is it the HGPIFuNetwNML model with HRI, Depth, Normal and Parse Map deactivated? 3) How many epochs was the vanilla PIFu model trained for? 4) Is spatial sampling or Depth Oriented Sampling used in training PIFu?
Thank you!