search

bearpaw / PyraNet

Code for "Learning Feature Pyramids for Human Pose Estimation" (ICCV 2017)
Apache License 2.0
221 stars 58 forks source link

How to generate 'scale' and 'objpos'? #15

Open Binyr opened 5 years ago

Binyr commented 5 years ago

Hi, Thanks for your code! The 'scale' and 'objpos' is also not provided offictially. How do you generate them? Expect for your guidance! Thank you!

Binyr commented 5 years ago

I read your paper again. 'objpos' is image center. But I still do not know how to generate 'scale'.

bearpaw commented 5 years ago

You can find the code to generate the JSON file here: https://github.com/bearpaw/pytorch-pose/blob/master/miscs/gen_lsp.m

Actually, it is borrowed from Convolutional Pose Machine: https://github.com/shihenw/convolutional-pose-machines-release/blob/a345bc5c760073792b6b5ec29d3bde01557a5a9b/training/genJSON.m

I didn't dig into the equations. But it should be some statistics from the training set (e.g., the average height of persons are 200px or something like that).

bearpaw commented 5 years ago

Another thing of CPM is that, it uses additional training data from LSP and LSP extended to train the model for the MPII dataset. In this case, we want to make sure that persons in different datasets should have roughly the same scale. That's how CPM computes the scales.

Binyr commented 5 years ago

Thanks for your reply! I am confused by this complicate computation. It seems that joints location is used to generate 'scale'. Could I use it directly for evaluate my model?

xiaoxin05 commented 3 years ago

@Binyr hi，Would like to ask the question about the scale to understand？