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

Results for dance-twirl #3

Closed iris112358 closed 2 years ago

iris112358 commented 2 years ago

Hi, Thanks for open-sourcing this awesome work. I tried to run on the dance-twirl video and got the following results:

https://user-images.githubusercontent.com/96632902/152095343-c2fc9026-7835-468a-be92-a2114da35bf1.mp4

It seems the rendered frame does not match the input frame, I organized the data as suggested in the Readme and run by the command sh scripts/dance-twirl.sh and test by bash scripts/render_result.sh dance-twirl log/dance-twirl-1003-ft2/pred_net_latest.pth --catemodel --cnnpp `, could you please let me know if this result is expected or if there's something wrong? Thanks!

gengshan-y commented 2 years ago

The args passed to the bash script seems wrong. Can you run 

bash scripts/render_result.sh dance-twirl log/dance-twirl-1003-ft2/pred_net_latest.pth 36

instead?

iris112358 commented 2 years ago

It works now. Thanks! BTW, is it possible to use an existing template for this model?

gengshan-y commented 2 years ago

The sphere template is loaded here. You can replace it with another template mesh.

You might need to tune other parameters to make it work, e.g., freezing the shape and only optimize cameras for the first few epochs.

iris112358 commented 2 years ago

The sphere template is loaded here. You can replace it with another template mesh.

You might need to tune other parameters to make it work, e.g., freezing the shape and only optimize cameras for the first few epochs.

I see. Thanks!