Question about evaluation on 3DPW

[mimiliaogo]

Hi,
I noticed that in your code, you evaluate the final 3D vertices results (MPVPE, MPJPE) after adding predicted camera parameters. Of course, the gt camera params are also added to the gt smpl vertices before evaluation. However, I notice in other papers ( like METRO), the 3D vertices are evaluated without camera predictions, which means the accuracy of camera predictions will not affect the final results.
Are these differences in the evaluation process make it incomparable between you and others?
Thanks.

[hongsukchoi]

Hi, I don’t understand how camera parameters are used in the evaluation 2022년 11월 25일 (금) 오전 10:10, Mimi Liao ***@***.***>님이 작성:

Hi, I noticed that in your code, you evaluate the final 3D vertices results (MPVPE, MPJPE) after adding predicted camera parameters. Of course, the gt camera params are also added to the gt smpl vertices before evaluation. However, I notice in other papers ( like METRO), the 3D vertices are evaluated without camera predictions, which means the accuracy of camera predictions will not affect the final results. Are these differences in the evaluation process make it incomparable between you and others? Thanks. — Reply to this email directly, view it on GitHub <#56>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AHNGKWYORMSTAZLCZL34RE3WKAGZPANCNFSM6AAAAAASK3ANUY> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

-- Best regards, Hongsuk Choi https://hongsukchoi.github.io

[mimiliaogo]

In

Pose2Mesh_RELEASE/data/PW3D/dataset.py

Line 93 in 7f24836

smpl_mesh_coord, smpl_joint_coord = self.mesh_model.layer[gender](smpl_pose, smpl_shape, smpl_trans)

, the ground truth 3DPW vertices are generated with camera trans.

[hongsukchoi]

Hi, GT should be in the camera coordinate system. Other methods assuming a single image input will do the same. Where are the predicted camera parameters used? 2022년 11월 26일 (토) 오전 1:13, Mimi Liao ***@***.***>님이 작성:

In https://github.com/hongsukchoi/Pose2Mesh_RELEASE/blob/7f24836c36dfdd52be6735505f44af11ec97e666/data/PW3D/dataset.py#L93, the ground truth 3DPW vertices are generated with camera trans. — Reply to this email directly, view it on GitHub <#56 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AHNGKWYI6ZDOFCJ5MONJUPTWKDQUHANCNFSM6AAAAAASK3ANUY> . You are receiving this because you commented.Message ID: ***@***.***>

-- Best regards, Hongsuk Choi https://hongsukchoi.github.io

[mimiliaogo]

@hongsukchoi
Sorry, I found that the predicted camera parameter I mentioned is in your another repo:
https://github.com/hongsukchoi/3DCrowdNet_RELEASE/blob/6e773064c8d6950b382f66d76b615aada4f2594b/main/model.py#L65
Also, what do you mean in this sentence: "Other methods assuming a
single image input will do the same?"
Doesn't everybody use the same 3DPW dataset?
Thank you so much!

[hongsukchoi]

Hi, Do you know the camera coordinate system, the world coordinate system, and the smpl coordinate system? Best regards, Hongsuk Choi https://hongsukchoi.github.io/ 2022년 11월 28일 (월) 오전 1:19, Mimi Liao ***@***.***>님이 작성:

…

@hongsukchoi <https://github.com/hongsukchoi> Sorry, I found that the predicted camera parameter I mentioned is in your another repo: https://github.com/hongsukchoi/3DCrowdNet_RELEASE/blob/6e773064c8d6950b382f66d76b615aada4f2594b/main/model.py#L65 Also, what do you mean in this sentence: "Other methods assuming a single image input will do the same?" Doesn't everybody use the same 3DPW dataset? Thank you so much! — Reply to this email directly, view it on GitHub <#56 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AHNGKW2L2ESYUICCTQJU6RLWKOCZXANCNFSM6AAAAAASK3ANUY> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[mimiliaogo]

Hi,
In my understanding, there are world coordinates (3D), camera coordinates (3D), and pixel coordinates (2D).
And the model's output is already in camera coordinates, so the predicted camera is for the human mesh projecting from 3D camera coordinates to 2D coordinates.
Therefore, I think the predicted 3D mesh has nothing to do with camera params. Camera parameters are used for 3D->2D projection.
This is also mentioned by FastMETRO author in postech-ami/FastMETRO#3 (comment)
BTW, can you explain a little about smpl coordinate? I'm not sure what I think is the same as you.

Thank you so much.

[hongsukchoi]

To conclude, you can compare Pose2Mesh with MeTro or any other methods. Parsing on GT 3D meshes is
just transforming the world coordinate parameters to the camera coordinate system.

The SMPL coordinate system is the world coordinate system, where the template SMPL mesh lies.