search

dumyy / handpose

CrossInfoNet of CVPR 2019 for hand pose estimation
128 stars 26 forks source link

MSRA crop #3

Closed cheolhwanyoo closed 5 years ago

cheolhwanyoo commented 5 years ago

Thank you very much for sharing your project!

When I trained msra using depth thresholding or pretrained com from v2v-posenet, performance was not good as I expected .

I want to know how you get the com in msra dataset.

Thank you.

dumyy commented 5 years ago

@cheolhwanyoo I have updated the MRSA train and test code. There are some tips, hope that can help you.

  1. MSRA dataset has 2 versions. We used the order while the v2v-posenet used the new. So we didn't use its com. Here we use a com （center of mass） calculated by the depth map samed as deep-prior.
  2. In the data preprocessing, we use a dynamic bounding box to crop the training dataset , while testing, we use a rigid size of (175,175,175) to deal with different subject's hand sizes. You can find them in the data/importers.py/
  3. For the online data augment, change the camera info (fx,fy,cx,cz) .
cheolhwanyoo commented 5 years ago

Thank you very much for your kind answers!! I'll try it again!