Yu Sun

It seems that `torch.distributed.launch` has been dropped in new version of Pytorch. In latest version, they use the `torchrun` instead. I have tested that this will work ``` CUDA_VISIBLE_DEVICES=${GPUS} nohup...

This is pretty weird. You don't have this basic python package? please try import types

Thanks for your attension. If you can obtain the tracking results of each person, you can take advantage of [OneEuro Filter](https://github.com/Arthur151/ROMP/blob/48b8dde4df507f4c15b0c0fd3c0144bb3f0e0bf0/src/lib/utils/demo_utils.py#L113) to smooth the mesh.

For ROMP, each estimated body center location is pointed to each person. The body center is supposed to be the center of torso (L/R shoulders and L/R hips). You can...

Good question. If we want to optimize the temporal smoothness, we have to know the tracking results of multiple people. While currently released code lacks this part. We are developing...

We have released the training code.

Thank you for your interest in our work. We have prepared a commit of training code. However, ROMP is under submission. We are still waiting for the final decision to...

The ground truth SMPL parameters are 1) parsed from MoCap data using MoSh algorithm; such as Human3.6M dataset 2) downloaded from [EFT](https://github.com/facebookresearch/eft)

ROMP doesn't supervise the camera parameters. What we supervise is the projected 2D joints, which is calculated using the estimated camera parameters + 3D pose. Learning 2D joints would help...

Yes, we freeze the SMPL parameters during training. The statistical parameter of SMPL is not supposed to be changed during training.