xinghaochen/Pose-REN

About Data Layer

liuxunfu opened this issue · 4 comments

liuxunfu commented

Hi,thanks for sharing your project, I have two questions here：
1 Is the training data layer defined in python?
2 In the training label of icvl data set, there are 48 floating-point numbers. How to realize multi-label training?

xinghaochen commented

Hi，

  1. You can implement the data layer in either Python or C++, we used C++.
  2. The pose label is a vector of size (48, ). Therefore, just feed the annotations as a blob of size (batch_size, 48) into the model.
liuxunfu commented

Hi，

1. You can implement the data layer in either Python or C++, we used C++.

2. The pose label is a vector of size (48, ). Therefore, just feed the annotations as a blob of size (batch_size, 48) into the model.

Ok, thank you. By the way, how is the following achieved?
"We extracted a fix-sized cube from the input depth image. The center of the cube was determined by calculating the centroid of mass of the hand region. The extracted cube was then resized into a patch with size of 96 × 96 and the depth values within it were normalized into [−1, 1]."

liuxunfu commented

Please see src/utils/util.py and src/utils/model_pose_ren.py.

Ok, thank you very much.