How you produce 'maskcoefficients' about yolact HOT 5 CLOSED

A 1x1 kernel conv layer is just a fc layer applied individually to each pixel in the image. That's what I meant by saying "as the product of an fc layer". One conv layer is producing mask coefficient for every pixel in the conv out individually--doing it with a conv is just way more efficient.

from yolact.

@dbolya from the paper it seems that the mask coefficients are linear , is their any way we can make the coefficients non linear . If so how could we approach .

from yolact.

@abhigoku10 It's not exactly linear depending on which place you look. The coefficients themselves go through a tanh, so they're all from -1 to 1. Then once they get multiplied with the prototypes (which are ReLU'd so they go from 0 to inf), that whole thing goes through a sigmoid. You could add some coefficients that get multiplied into a square prototype or something like that (similar to geometric [nonlinear] regression).

from yolact.

@dbolya i wanted to improve the accuracy for person detection in mscoco and custom data for that when i visualisation the protomask i am getting the features but the mask coefficients is low so increase that

from yolact.

A 1x1 kernel conv layer is just a fc layer applied individually to each pixel in the image. That's what I meant by saying "as the product of an fc layer". One conv layer is producing mask coefficient for every pixel in the conv out individually--doing it with a conv is just way more efficient.

Hi, in the PredictionModule notes, you said 'this is slightly different to the module in the paper', is this meanings change the extra_head_net kernel size from 11 to 33, and is this original 1*1 the fc layers mentioned in your paper? And now there is no fc layer in prediction of mask_coefficients?

from yolact.