Question about training ICON with custom scans
Closed this issue · 2 comments
Hi, sorry to interrupt you but I have question about training ICON with custom dataset.
I have custom scans and SMPL parameters fitted with them(with this open source), and tried to generate train data for ICON, but found that your data generation code only accepts SMPL-X parameters that contain left/right hand poses and facial poses in pkl file, according to lib/renderer/mesh.py code.
In issue #199 you commented to try using ECON for making data from scans, but I'm not sure how much reliable ECON is, since it only uses single image to estimate SMPL-X parameters.
I was curious how did you get SMPL-X fits for other datasets like RenderPeople, which doesn't have SMPL-X fits on its own, when you were training ICON for the paper? Did you just make pseudo SMPL-X fits with other open sources that estimates SMPL-X parameters with single image? Do they have enough quality to train models like ICON?
You can find the SMPL-X fits of RenderPeople from AGORA dataset.
Same issue: YuliangXiu/ECON#62
Our internal fitting pipeline is similar to MultiviewSMPLifyX, but there is no plan to release.
As for the poses of hand and face, you can edit the class of SMPLX in ICON, to make it compatible with your setting. ICON supports any parametric body models as long as these models could produce mesh-based body as input.
Thank you for quick respond! I'll try editting the SMPLX class, and share if it works :D