About the training data of example model #46
-
|
Hello Nicholas, In continuation with our previous discussion, I really wonder which data and what setting do you use to train the example_model.th in link. Because I tried to use the pretrained model model.th provided in link, which is supposed to be trained on synthetic Human3.6M, correct? Unfortunately, using this model, I can't get the same almost-perfect predictions on U3DPW as those fitted by the inference model example_model.th ... Please do help me clarify this, as it is quite important to know which model is trained on what, for the sake of a fair comparison. I did appreciate your answers to my previous questions! Thanks a lot, |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 1 reply
-
|
Hi, both are trained with synthetic data. The pre-trained model S15678 was trained with synthetic data and norm_single=False. This model is intended for in-door evaluation on H36M. The inference model (example_model.th) was trained with synthetic data and with norm_single=True. Note that the normalization method must match for both training and inference.
|
Beta Was this translation helpful? Give feedback.
-
|
For reasons why two normalization methods are used, please see #12 (comment). For in-the-wild images, the location of the subjects carry little information of their 3D poses. Thus it is OK to make the input translational-invariant. For H36M, the default normalization leads to better performance. This means that subject locations are useful for inferring their 3D poses in H36M (e.g., a person is at a particular location is informative for inferring its pose). However, this is actually dataset bias and does not generalize to unconstrained images.
|
Beta Was this translation helpful? Give feedback.
-
|
Hello Nicholas, Sorry for a late response! I will try them and let you know, thanks a lot for those information! Best, |
Beta Was this translation helpful? Give feedback.
Hi, both are trained with synthetic data.
The pre-trained model S15678 was trained with synthetic data and norm_single=False. This model is intended for in-door evaluation on H36M.
The inference model (example_model.th) was trained with synthetic data and with norm_single=True. Note that the normalization method must match for both training and inference.