/sci/ - Science & Math

[Anonymous No.12352335]

I will pay $20 to whoever can help me figure out this fucking paper
https://openaccess.thecvf.com/content_CVPR_2020/html/Hou_Learning_to_Structure_an_Image_With_Few_Colors_CVPR_2020_paper.html

I'm a fucking brainlet, please help.
The topic is convolutional neural network and color quantization
I'm can understand the first 3 equations (the third one kind of) , but after that, I can't manage.

I'm confused on T(x) speficially (why its shape is C x 3), which depends on M(x), but I don't know what a color index map exactly is (although I have some idea).

[Anonymous No.12352378]

>>12352335

Problem:
If the color space of an image consist of a large patch of the same color certain quantization algorithms tend to make the image unrecognizable for the AI. The rest is really just showcasing other techniques and algorithms that get around this and comparing them. Have fun.

[Anonymous No.12352416]

>>12352378
I understand the goal of the paper, and what it tries to achieve
I'm just confused on the exact details on how they achieve that.
They claim that unlike other color quantization methods, ColorCNN looks at structure and color instead of just color, like the MedianCut or OCTree would do.

They just say the use an "Auto-encoder" to identify these so called structures, but that isn't explained in the paper.

I have to give an hour presentation on this very soon, I'd like to understand as much of it as possible.

[Anonymous No.12352737]

bump