Difference between inference code and evaluation code?
Closed this issue · 1 comments
Hi authors, I noticed a difference in the predicting mesh using inference and evaluation code.
Specifically, for THuman 2.0 human id 506, the predicted meshes from running inference code for the PiFU model
python -m apps.infer -cfg ./configs/pifu.yaml -gpu 0 -in_dir data/thuman2_36views/0506/render
and testing code
python -m apps.train -cfg ./configs/train/pifu.yaml -test
are different.
See in the below image. Left is inference code. Right is testing code. The input image have the same input view.
Input image
Could you please help clarify why this is the case? Thank you in advance!
In the inference mode, the SMPL-X body is estimated using off-the-shelf methods, such as PIXIE or PyMAF.
However, in the evaluation mode, we use ground-truth SMPL-X body to filter out the influence of body misalignment. Thus, the final reported numbers could solely reflect the reconstruction performance of clothing instead of the body pose.