Still vastly underused in image to image to this day and I don't understand why: Unets and self-attention. I keep seeing papers come out that have obvious problems that could be solved using these two simple things!
-
I'm curious to see what U-Nets with EfficientNet backbone can do. I was also reading the DeepLabv3 paper, it seems that its dilated convolutions are wonderful to preserve fine grained details. It looks like U-Nets win doing blob-like segmentation while DeepLabv3 wins on details.
1/ The only EfficientNets I've had some success with are B3 and B4- but still not as good as ResNet. You get all excited about using a backbone with a big increase in accuracy, but it seems in practice that these nets are highly optimized in one single direction- visual rec.
