search

facebookresearch / VideoPose3D

Efficient 3D human pose estimation in video using 2D keypoint trajectories
Other
3.65k stars 746 forks source link

About evaluation on datasets #18

Open DongJT1996 opened 5 years ago

DongJT1996 commented 5 years ago

Hello, I have one problem about your paper, you only evaluate on Human3.6M, HumanEva, which I think is not enough to prove the generalization ability of the model. Previous papers usually evaluate on the dataset MPI-inf-3DHP. Have you evaluated on it? Otherwise, the model may be just overfitting the dataset Human3.6M.

michaelauli commented 5 years ago

As you say, our method works on two datasets, and therefore is not overfitting to one particular dataset.

DongJT1996 commented 5 years ago

Thanks for your reply.

What I understand is that you use the Human3.6M data to train a model, and test the model on the Human3.6M, use HumanEva to train another model, and test on the HumanEva. However, previous papers usually evaluate their model on the dataset MPI-inf-3DHP without using training data from the dataset, which is just for testing.