pixelite1201
commented
1 year ago Hi Fabian, Thanks! Yes we have already released SMPL-X parameters for BEDLAM. We will also be releasing SMPL parameters generated by fitting.
During training we always predict SMPL-X body pose parameters in BEDLAM-CLIFF and BEDLAM-HMR and the loss is also on SMPL-X pose parameters since all the ground truth labels used in training (BEDLAM and AGORA) are in SMPL-X format. We never use H36M, MPI-INF-3DHP etc. in training BEDLAM-CLIFF or BEDLAM-HMR. For evaluation on 3DPW and other SMPL based datasets, we use SMPL-X to SMPL vertex regressor on our predictions to calculate the errors.
If you are talking about the CLIFF model train on real images like H36M, MPI-INF-3DHP, we use only SMPL based ground truth there provided by respective datasets so all the losses are on SMPL pose parameters. I hope it is clear. Let me know if you have more questions.
fabienbaradel
Hi, First of all congrats for accepted paper to CVPR. Thanks a lot for sharing the code and the dataset. This will be very helpful for the research community working on this field. As far as I understand you are releasing SMPL-X parameters for BEDLAM. You are also releasing the regressor for mapping the SMPL-X vertices to the SMPL vertices. It seems that you are using it when you are training HMR for Human Mesh Recovery for estimating SMPL params. But you are computing the losses on the SMPL pose parameters between the ground-truth SMPL-X pose parameters and the predicted ones given by your model. When using multiple training sets (e.g. H36M, MPI-INF-3DHP) you only have SMPL pose parameters, so it is okay if there is "mismatch" between SMPL-X and SMPL pose params? I hope that my question is clear enough, Fabien