Open HongwenZhang opened 1 year ago
Hi Hongwen
Thanks for letting me know.
Considering the compatibility, I pre-processed the input image with my own implementation (crop parts w/ mediapipe) and used PyMAF-X's ckpt for inference. If I want to integrate the updated version, should I change the pre-processing steps (cropping, transformation), or just need to update the ckpt, from v1.0 to v1.1?
It should be fine to use your own cropping implementation and simply feed the body/face/hand part images to PyMAF-X. As there are several modifications on the face-specify regression network of the updated PyMAF-X, the ckpt v1.1 is no longer compatible with the original one. It is recommended to replace all configuration and model definition files with their updated version.
Other minor issues:
i) set MODEL.PyMAF.OPT_HEAD = True
if the head orientation is preferred to be exactly identical to the face-specify prediction result (see here). If OPT_HEAD is False, the head orientation is the same as the body-specify prediction.
ii) MODEL.PyMAF.HF_BOX_ALIGN
is set to False
in v1.1 but True
in v1.0 by default. This change should help to have slightly better face/hand regression.
Congrats to ECON! 🎉🎉
It is just a kind reminder that an updated version of PyMAF-X is now available for better face and hand mesh regression. i) the updated PyMAF-X uses the same version of SMPL-X (v2020) as PIXIE; ii) the face and hand inputs of PyMAF-X are simplified (no need to provide local transformation information); iii) the face regression performance should be on-par with PIXIE, see Fig. 7 and Table 4 of the updated paper;
Please check the updated PyMAF-X for more details :)