I'm learning more about different variations of deep CNNs. Based from my understanding, ResNet makes use of skip connections that's also somehow shaped like a pyramid or triangle? How is this different to Feature Pyramid Network then?
Typically people use FPNs for segmentation/detection tasks (dense pixel-wise predictions), with skip connection between different levels^1 of the network, while ResNet is used for classification/regression tasks (sparse, image-wise predictions), with skip connections within a residual block (between different layers of the network).
Another difference is, typically, in FPN you do perform some kind of upsample of feature maps (in a form of deconvolution or a interpolation followed by convolution), while ResNet consist of sequence of downsampling transformations (convolutions and poolings).
Schematically, FPNs could consist of ResNet (which transform spatial width x height into features), followed by a "reversed ResNet" (which transforms features into spatial width x height), with skip-connections between corresponding spatial levels, forming a pyramid, hence the name.