could you share the codes for training/testing the MPI-INF-3DHP dataset?

[luzzou]

Hello~ I am recently testing the generalization ability of Human3.6M-trained models on the MPI-INF-3DHP dataset, but I always failed to get the results as the paper demonstrated. Could you share your codes for training or directly testing the MPI-INF-3DHP dataset? or releasing your pretrained model on this dataset? Some data pre-processing steps can also help me a lot!

Thanks in advance! I'm looking forward to your reply

[AminAnsarian]

Hi! Did you get to solve your issue in this matter? I also have some trouble running on the mpi-inf-3dhp dataset, mainly on the preprocessing step. I'd appreciate your help if you can.

Thanks in advance!

[sunnychencool]

Hi Amin,

Thank you for pointing out the preprocessing step. Indeed, I have graduated and left the original position, so I'm not holding the code now. However, I can share with you the key points. For the MPI-INF-3DHP dataset, I extract each training/testing frame's 17 keypoint's 2D positions and their visibility confidences by AlphaPose. It should be noticed that the keypoints predicted by AlphaPose are in the COCO format (i.e., 12 limb keypoints + 5 face keypoints), and they are the input of our proposed 3D model to predict the 14-joint skeleton defined by MPI-INF-3DHP. Another significant thing is that MPI-INF-3DHP's training data contain a great number of frames where the actors are not in the view of the camera (typically, they are at the end of each video). A simple filter should be used to filter out all these invalid training frames based on the ground-truth 2D keypoints (in this situation, the ground-truth 2D keypoint positions will be out of the range of the frame image's resolution). This operation will remarkably improve the model's performance. The other parts completely correspond to our released code. The 3D bone definition part of the code should be modified to match the 14-joint skeleton of MPI-INF-3DHP.

From: Amin Ansarian @.> Sent: 2021年4月8日 13:49 To: sunnychencool/Anatomy3D @.> Cc: Subscribed @.***> Subject: [EXT] Re: [sunnychencool/Anatomy3D] could you share the codes for training/testing the MPI-INF-3DHP dataset? (#5)

Hi! Did you get to solve your issue in this matter? I also have some trouble running on the mpi-inf-3dhp dataset, mainly on the preprocessing step. I'd appreciate your help if you can.

Thanks in advance!

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_sunnychencool_Anatomy3D_issues_5-23issuecomment-2D816181142&d=DwMCaQ&c=kbmfwr1Yojg42sGEpaQh5ofMHBeTl9EI2eaqQZhHbOU&r=rPOBJC5gKaoaPCPUTVTptPi9X8fSUVufMYyYR7i8ztA&m=eMSBOeXGYtuXjqbgIDLA2lBkWpwD0cqjRa2PEo3qyfM&s=alBOqwZOyvhDmCJuSZL_c4T8jflEehNKryTd76-P2os&e=, or unsubscribehttps://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_notifications_unsubscribe-2Dauth_AFQKPEHJVHL3YUHYVIMX7HLTHYJC5ANCNFSM4P57VIEQ&d=DwMCaQ&c=kbmfwr1Yojg42sGEpaQh5ofMHBeTl9EI2eaqQZhHbOU&r=rPOBJC5gKaoaPCPUTVTptPi9X8fSUVufMYyYR7i8ztA&m=eMSBOeXGYtuXjqbgIDLA2lBkWpwD0cqjRa2PEo3qyfM&s=Do_WHATi4nSUIDA7A-vSdbnam64MmeNMgq2_8Tdw4V4&e=.

[AminAnsarian]

@sunnychencool Thank you very much for your response. I appreciate the detail and comprehensiveness.

However, everything up until the 3D training is complete. As you may know, the VideoPose3D has a complex multistep preprocessing based on the camera parameters of the Human3.6m dataset, including the extrinsic and intrinsic parameters. This, however, is very different in the case of the mpi-inf-3dhp dataset. So my question is, if I may, right after collecting all of the extracted 2D joints. How exactly do you feed them to the VideoPose3D and by extension, your model.

Looking forward to your reply,

Sincerely, Amin

[hcyz33]

I can share with you the key points.

hello, could you share me the keypoints of mpi, including 2D keypoints from coco and 3D joints after seleted from annot3d. thank you every match!

[nies14]

did u get those? if yes, could u please share it with me? mthnies@gmail.com

[CapnBloodBeard12345]

@nies14 @luzou-ustc @hcyz33 did you guys have any luck figuring out what to modify in the randomaug method of generators.py to train on the MPI dataset?