search

gengshan-y / viser

ViSER: Video-Specific Surface Embeddings for Articulated 3D Shape Reconstruction. NeurIPS 2021.
https://viser-shape.github.io/
Apache License 2.0
73 stars 6 forks source link

Err result with ama-female dataset #10

Open bravotty opened 2 years ago

bravotty commented 2 years ago

I have a problem in reproducing the ama-female dataset from viser-rlease. I used the init-cam provided by Posenet, I also used your updated ama-female.sh script, and I used the banmo format in vid.py, and changed the smooth loss to the original 0.005. In the end the training of ama-female resulted in an image like this, what could have gone wrong with this process?

4786b8ea38163371c5191b6bf3f0b86

bravotty commented 2 years ago

Similarly, I also appear to have a segmented human body, how can I get the situation in your paper? e9f3226edcaf7a4dc1ff8f309ba660e

gengshan-y commented 1 year ago

This looks extremely strange. I suspect some non-algorithmic errors exist. You want to verify the code works for the breakdance to make sure all packages are installed correctly.

Then I need more information to look into this, for instance, the tensorboard logs (scalar + images) of the first stage of training.