Why do I need to load 2D data when training 3D data？

[APeiZou]

Why do I need to load 2D MPII dataset when training the 3D Human36M dataset？

[mks0601]

The model is jointly trained on 2D MPII and 3D H36M datasets.

[APeiZou]

@mks0601 They are have deffierent num_keypoints? how many outputs numkeypoints?

[mks0601]

The joints are transformed to those of 3D datasets (https://github.com/mks0601/3DMPPE_POSENET_RELEASE/blob/3f92ebaef214a0eb1574b7265e836456fbf3508a/data/dataset.py#L105)

[APeiZou]

@mks0601 Can I only train in Human36 datasets ? How to init_weight of the model when training? Do I need to load the pretrained model?

[mks0601]

You can teain only on h36m by removing mpii from the traingset. You dont have to load it.

[APeiZou]

The values of vis are all zero when I load Human36 dataset，and the loss is zero。

[mks0601]

Please give me full descriptions of your situation. I cannot give you an answer from a screenshot.

[APeiZou]

The vis Attributes of target is zero and the loss are always zero。

[mks0601]

joint_vis is set in here. The original codes work well, so you'd better check your modifications.

[APeiZou]

@mks0601 I have not bbox_root_human36m_output.json file ，can I train the model in Human36 datasets ?

[mks0601]

You can downlaod from here. https://github.com/mks0601/3DMPPE_POSENET_RELEASE#results

[st20042066]

@APeiZou Hi, did you fix the problem about the vis Attributes of target is all zero? I think I also have this problem and I don't know what's wrong. Can you give me some advice please? Thank you so much.

[st20042066]

@mks0601 I print the conditional parameter and find out that the data has huge error, Can you help me please? Thanks a lot!