Open leilaUEA opened 8 months ago
I have another question, about 2D detections in the 3DPW dataset. In the original dataset, there is 18 joints but the training dataset has 17 joints. Are these 17 moving joints of human 3.6m? Are they computed using OpenPose, Stacked hourglass or re-projection of 3D pose?
In the paper you mentioned: "We do not use the images or 2D detection results of the two datasets during pretraining as Mocap datasets usually do not provide raw videos. Instead, we use orthographic projection to get the uncorrupted 2D skeletons"
I have a question about training the model using different data. Assuming having the SMPL ground-truth and the corresponding videos/images, Is there any code to create the training data in a format readable by the mesh training code? If I can create a similar .pkl file for the new custom dataset that would be great.
Thank you very much in advance, Looking forward to hearing from you