Hand parts result on InterHand26M

[aragakiyui611]

I tested the hand part of Hand4Whole and the result is MPJPE: 63.86 mm, is it reasonable? Did you test the model in terms of MPJPE on InterHand26M and may I know your result? Thank you!

[mks0601]

For the hand-only images, you should use Pose2Pose (https://github.com/mks0601/Hand4Whole_RELEASE/tree/Pose2Pose)

[aragakiyui611]

Yes I am using the pose2pose

[mks0601]

mpjpe 63 is not a making sense number. it's too big. I guess there should be some bugs? How did you evaluate on ih2.6m dataset?

[aragakiyui611]

InterHand26M.py.txt model.py.txt I using the same code as that in the repository, my trained weights perform normally on other datasets like FreiHAND, COCO

[mks0601]

joint_gt includes both right/left hands, while the output only includes the right hand.

[aragakiyui611]

Thank you! I flip back the left hands and MPJPE now is 18.21mm

[mks0601]

Please note that for the fair comparison to previous works, you should train the network again after removing this view sampling. https://github.com/mks0601/Hand4Whole_RELEASE/blob/39fd3727ac1729c0ce59890986e41abbce0f9955/data/InterHand26M/InterHand26M.py#L56

[aragakiyui611]

Thank you!

[aragakiyui611]

Hi, I want to download the Human36M dataset you shared, but the google drive said it exceeds the download quota. "此文件已超出下载配额，因此目前无法下载"： this file have exceeded download quota, thus cannot be downloaded now.

[mks0601]

Try this at below

https://github.com/mks0601/I2L-MeshNet_RELEASE#data

[aragakiyui611]

Try this at below

https://github.com/mks0601/I2L-MeshNet_RELEASE#data

It seems that this is the one I use, I cannot download it either

[mks0601]

Try a trick written below

[aragakiyui611]

Thank you every much!