zhizdev
Doubts about depth scale of the provided GSO test data
Hello author, according to the provided blender scripts, the original rescale should be [0,1]->[0.5,2.5], but I use the following code for the backprojection:
depth_img = depth_img * 2.0 + 0.5 # !
depth_img[depth_img == 2.5] = 0
There are some noticeable misalignment at the edges in the final projected pointcloud:
I follow the depth process in Syncdreamer and find its depth scale is [0.6,2.4]: https://github.com/liuyuan-pal/SyncDreamer/issues/50
With the Syncdreamer's rescale the projected pointcloud is better aligned:
I would like to ask if there is a problem with my findings, could it be that the training of mvdfusion is rescaled to [0.5,2.5] but the GSO used is rescaled to [0.6,2.4]?
Hi @YZsZY thanks for checking this out and posting the findings!
We don't have enough information to comment on your findings atm but one thing that is true is that MVD-Fusion is trained on Objaverse with [0.5, 2.5].
