Evaluating on 3DPW and MuPoTS

[gurkirt]

I am new to pose estimation and all the dataset with different skeleton models are making me confused.

If I train on the Human3.6M dataset, which all dataset can I evaluate? Can I evaluate on both MuPoTS and 3DPW? I see that your RootNet does work for both these datasets but not PoseNet, is there a specific reason? Similarly, I2L-MeshNet is evaluated on 3DPW but not on MuPoTS?

Also, mostly, people use the MuCo dataset instead of the MPI-INF-3DHP. I understand MuCo is an augmented version of MPI-INF-3DHP, which has backgrounds of MSCOCO. However, why not use both MPI-INF-3DHP and MuCo?

I want to estimate 3D poses in videos but not meshes. However, I want the ability to evaluate both MuPoTS and 3DPW while trained in Human3.6M. Should I use your I2L-MeshNet repo or PoseNet repo?

Where can I find skeleton models of each 3D dataset that is publically available?

I would be grateful for your help.

Many thanks Gurkirt

[mks0601]

There are too many questions... I'm too busy to read and answer all your questions. Please summarize the questions and itemize them.

[gurkirt]

Thank you for your reply.

sure, here it is.

Here is my premise first up, I want to estimate 3D poses in videos but not meshes. However, I want the ability to evaluate both MuPoTS and 3DPW while trained in Human3.6M.

Questions

1. Can I evaluate on both MuPoTS and 3DPW if I train PoseNet only on Human3.6M?
2. if not, why?
3. Should I use your I2L-MeshNet repo or PoseNet repo for the above task? as you know, 3DPW is not there in posenet but in I2L-meshnet setup.

These three will help greatly.

[mks0601]

1. You can evaluate as there are common joints in those datasets. However, I think the performance would be bad as images in Human3.6M are captured only from multi-view studio. On the other hand, images in MuPoTS and 3DPW are captured in outdoor environment.

2. See 1.

3. Both I2L and PoseNet will work.

[IemProg]

Hi,

I'm wondering about the cropped image fed to the model after extracting the person for the MuCo dataset. I have noticed that the given size is so big [width / 4, height / 4], it contains many persons in the image which can be very noisy for the model (since persons in some scenes are superposed on each other ).

Have you considered changing the size of the cropped image fed to PoseNet or the bounding box dimensions given by COCO?

Thanks!

[mks0601]

Only a single person closest to the camera is considered during training on MuCo dataset. Therefore, I think there would be no big problem.