syguan96/boa

[CVPR 2021] Bilevel Online Adaptation for Human Mesh Reconstruction

Hi Coder 👋

My name is Shanyan Guan. I'm a Ph.D. student of SJTU, supervised by Professor Xiaokang Yang and Professor Yunbo Wang.

My research path is:
image generation --> 3D generation --> human mesh recovery --> StyleGAN inversion --> personalized generation ---> future

My profile overview:

boa's Issues

About reported performance on MPI-INF-3DHP in paper

In the Table 4 of BOA paper, it is reported that BOA achieves PA-MPJPE of 77.4. However, when I run the MPI-INF-3DHP code from this repository, I get PA-MPJPE of 92.61. Is there reason for such difference in performance between main paper and this code?

about 3DHP dataset

About 3DHP dataset, did you use the SMPL data provided by SPIN (mpi-inf-3dhp_mview_fits.npz) or the final fits data (http://visiondata.cis.upenn.edu/spin/spin_fits.tar.gz)

SMPL+H

Hi. Thanks for your great work. I wonder where to download SMPL+H model?

Time and device to train the network?

Hello,

May I know what GPUs do you use and how long does it take to train the model?
Your reply would be much appreciated. Looking forward.

Thank you!

inference time

Hello,

Thank you for your great work and for releasing your code. I have some questions about the procedure of test-time adaptation.

1. Do we need to fine-tune the network on each inference sample and then perform inference or we do fine-tuning on the whole test set and then at the end we perform inference.
2. How much is the inference time?

Thank you

can not download pretrain model through the provide website