Last week I finish the function for the pose skeleton normalization. Given a referenced skeleton expressed with global coordinates, we can take some joints of it(instead of all joints) and normalize the length of the chosen joints of the skeletons in other frames. All the computations are using PyTorch’s tensor, which can be easily used for Deep Learning, or transferred to Numpy array.
Also, to make the test work easier, I visualize the skeleton. I don’t find the existing implementation so I just do it myself.
In the figure, the blue one is the referenced skeleton. The red one is the original skeleton(useless). To see if the code works, I multiply all the coordinates of the red one by 2. which becomes yellow one. Then we normalize it using the length from blue one, and the green one is the result. The hip’s coordinate stays the same.