Doubts about training

[Vivien888]

Doubts about training

[xiexh20]

Hi,

We used 3 RTX8000 GPUs, each with 48GB memory to train the model. If you get the out of memory error, you can try to reduce the batch size.

No, we did not train on COCO and NTU-RGBD datasets as there are no 3D GT available for training. We only tested our method pretrained on BEHAVE on these datasets.

[Vivien888]

Thank you for your reply. I would also like to ask how can I test a single RGB image now. Do I have to generate "color.json " "mocap.[json|ply]" "_mask.jpg " files first and then input them with the RGB picture? If yes, how can I obtain the mask images? Looking forward to your reply.

[xiexh20]

Hi,

Yes, you need to prepare these files for your own input images.

To generate masks, you can use interactive segmentation such as this ritm, and for video you can also use video segmentation such as MiVOS.

The color.json file stores openpose predictions, you can simply run openpose on your images and save the results following the same format.

The mocap.json|ply stores the SMPL model parameters and SMPL meshes estimated by FrankMocap.

Hope that helps.