Interpreting `category_embedding`

[lmanan]

Hello @detectRecog , thank you for making your code public and for proving a very interesting approach to tracking. I have the same question as @ljjyxz123 in issue #16 .

I noticed that category_embedding is provided as a global parameter, how should we interpret this 4 x 3 float matrix?

In the publication, you mentioned that all semantic categories including the background are encoded into one-hot vectors. How does the provided category_embedding relate with this assertion?

Also, more specifically if one is training a tracker on grayscale (one channel) images, would the category_embedding change?

Thanks again for your code !