About the training stage to reimplement the reported results on 3dpw dataset

[zxk19981227]

I tried to write the training method by myself. However, i couldn't reproduce the results on 3dpw. Could you explain the training methods more detail? I tried to train the pymaf-x with the coco and itew provided in pymaf as reported in PARE, but after 15w step the results is not as well as reported in paper. Should i train the model with any other steps?

[HongwenZhang]

Hi, thanks for your interest in our work. Have you used the EFT labels? And how about the performance of the HMR baseline? It is also recommended to monitor the performance of 3DPW after every epoch.

[zxk19981227]

Thanks for answer, i use the eft labels provided in Pymaf to train the pymafx. I want to make sure that the models' performance on 3dpw is only traind on coco and other 2d keypoints dataset?

[HongwenZhang]

Hi, 3D mocap datasets such as Human3.6M and MPI-INF-3DHP are also needed.

[zxk19981227]

Hi, so the train have two stage. 1.train on 2d dataset 2.train on whole dataset. After these training, the model is evaluated on 3dpw?

[HongwenZhang]

Yes.

[zxk19981227]

Hi, I hope you could provide a more detailed reproduction plan. For example, first train on the COCO and MPII data sets, and then retrain the entire data 2d+3d data set on the best checkpoints. Although I can tell you Find their detailed training methods in PARE. But in your paper, I can hardly find any specific training methods after you join the MPII data set.

[HongwenZhang]

Thank you for your advice.