Codes for preprocessing MPI-INF-3DHP dataset

[nies14]

Could you share the codes for preprocessing the MPI-INF-3DHP dataset?

[Vegetebird]

I'm sorry for the late reply. I will release the code about 3DHP. You can also use other MPI-INF-3DHP code for preprocessing.

[nies14]

Can you give me some links to some repo that has MPI-INF-3DHP preprocessing code?

and which reorder index list u did use for preprocessing the MPI-INF-3DHP test dataset?

[Vegetebird]

Hi~You can refer to Anatomy3D issue #5, PoseAug, HMR, SPIN.

[nies14]

How many joints are used for testing MPI-INF-3DHP?

[alecda573]

@nies14 @Ha0Tang @Vegetebird anyone ever figure out how to process MPI and train this model on it as well as evaluate?

also does anyone know the protocol they evaluate this under

[nies14]

Hi @alecda573. You can check Ray3D or PoseAug

[alecda573]

@nies14 @Vegetebird neither of those provide codes or explain what their train/test split is.

have you found any codes for training or explaining the evaluation?

[nies14]

@alecda573

PoseAug tested their model on a cross dataset. They trained using Human3.6m and ground truth keypoints and then tested their model on the test set of MPI-INF-3DHP. You can use this script (https://github.com/jfzhang95/PoseAug/blob/main/run_evaluate.py) and to use the AUC and PCK metric remove comments from here (https://github.com/jfzhang95/PoseAug/blob/f3f5c4e916ebf7529b873ec1c14c1ce0bf0f5cb1/function_poseaug/model_pos_eval.py#L70)

But this MHFormer repo is not using the cross scenario evaluation.

[alecda573]

@nies14 ah good to know, so MhFormer does actually train on MPI and then evaluates on the validation set?

Have you used the code you just linked to try reproducing their results on MPI?

[nies14]

@alecda573

Yes.

I didn't reproduce the result of MHFormer. I used those codes to test mine.

[alecda573]

@nies14 are you currently setting up MPI to train MhFormer on?

have you been able to reproduce MhFormer results on CPN from scratch on H36M?

Would you be willing to share the evaluation Of MPI you got from training on H36M?

[nies14]

@alecda573 I'm not working on Mhformer and I didn't reproduce their results too.

I tested PoseAug and the results are almost similar what they mentioned in their paper.

[qiqiApink]

@nies14 PoseAug use 16 joints, but in MHFormer is 17. Did you find the preprocessing code that be used directly by MHFormer and has same format with h36m?

[nies14]

@qiqiApink No. I'm using cross dataset testing which is different from Mhformer. But according to the author you need to follow 14 joints skeleton structure for training & testing in mpi-inf-3dhp.

[anas-zafar]

@qiqiApink were you able to draft a script for cross evaluation scenario for Mhformer?

[qiqiApink]

@anas-zafar I used the code of P-STMO.

[anas-zafar]

Got it. Thanks @qiqiApink

[MehwishG]

@Vegetebird , I am trying to train MHformer on MPI-INF-3DHP dataset. I have followed the code of P-STMO but still getting a large MPJPE of 298. Can you please share the checkpoint for MPI dataset?

[GloryyrolG]

@alecda573

PoseAug tested their model on a cross dataset. They trained using Human3.6m and ground truth keypoints and then tested their model on the test set of MPI-INF-3DHP. You can use this script (https://github.com/jfzhang95/PoseAug/blob/main/run_evaluate.py) and to use the AUC and PCK metric remove comments from here (https://github.com/jfzhang95/PoseAug/blob/f3f5c4e916ebf7529b873ec1c14c1ce0bf0f5cb1/function_poseaug/model_pos_eval.py#L70)

But this MHFormer repo is not using the cross scenario evaluation.

Hi @nies14 , @Vegetebird , do you believe the results are actually trained on MPI-INF-3DHP instead of cross-dataset generalization? But the paper says it is cross-dataset generalization. Many thanks & regards:)

To assess the generalizationability, we evaluate our method on MPI-INF-3DHP dataset.