search

zlai0 / MAST

MAST: A Memory-Augmented Self-supervised Tracker (CVPR 2020)
https://zlai0.github.io/MAST/
273 stars 32 forks source link

Question about Lab channel dropout #19

Closed liyz15 closed 3 years ago

liyz15 commented 3 years ago

https://github.com/zlai0/MAST/blob/a57b043ca597b9b7ef6842b1fa965c9f1ee71526/models/mast.py#L43-L44

This is essentially

drop_ch_num = 1
drop_ch_ind = [np.random.choice([1, 2])]

In the paper all three channels of Lab are decorrelated. Is it on purpose that L channel is preserved?

Great work! Thanks in advance.

zlai0 commented 3 years ago

Hey sorry for the confusion. The paper actually says "One of the color channels is randomly dropped with probability p=0.5". Here in the context of Lab colour space, the colour channels typically refer to only ab, whereas the L channel is referred to as the lightness channel, which does not contain colour information. However I must say this line could be misinterpreted.

I looked at my original code and did find another version (i.e. all three channels could be dropped), but I forgot what the specific performance was. I assume the current version yields better performance..

liyz15 commented 3 years ago

Thanks for the clarification!