search

RenYurui / PIRender

The source code of the ICCV2021 paper "PIRenderer: Controllable Portrait Image Generation via Semantic Neural Rendering"
Other
506 stars 63 forks source link

Thank you for ur nice work. I have doubt about the 3DMM's translation parameters #28

Open summertight opened 1 year ago

summertight commented 1 year ago

I have thought that the translation parameters work as a translation of the rigid head in the original image coordinate(each x, y, z axis). However, the generally, when we get the 3DMM parameters from the 3DMM parameter encoder(Deep3Dpytorch), the head images are should be aligned from specific alignment algorithm to regular window. Then, the translation parameters are extracted from the that window, which has no relationship with the original image or the window position(before the alignment).

How the translation parameters work well on this model? I think that the position of alignment bounding box from the original image also must be considered with the translation parameter.

Thanks.

abm984 commented 1 year ago

hi I'm facing a problem regarding the PIRender setup. may you look for it in colab