Semantic Sermentation, about the mask?

Asked Jan 27 '18 at 08:33

Active Jan 27 '18 at 08:33

Viewed 76 times

If I get the mask on the feature map, for example, a 7 x 7 feature map, how can these 7 x 7 masks finally become the mask on the raw image? Using interpolation or upsampling? As far as I'm concerned, if you do mask on a 7 x 7 feature map where you only have 49 pixels, will it be too rough to mask the object in the raw image, for you cannot be too specific on only 49 pixels...

Thanks a lot!!

asked Jan 27 '18 at 08:33

YJHMITWEB

You simply use a bigger mask, where are you getting the 7x7 feature map from? – Dr. Snoopy Jan 27 '18 at 19:25
In Mask R-CNN, the author uses the resnet/fcn last layer feature which is 7 x 7 or 14 x 14, and he just does mask operation on them. – YJHMITWEB Jan 28 '18 at 07:22
1

Mask R-CNN outputs a 28x28 mask, not a 7x7 one. – Dr. Snoopy Jan 28 '18 at 10:22

Semantic Sermentation, about the mask?

0 Answers0