Feature Pyramid Network and RetinaNet

PhD Vlog talked about some OD networks, and this is the development line of ODs

0 FCN (Fully Connected Network)

This early work is barely used these days. Expand to full size image from the bottleneck loses lots of informations U-Net was more widely used b/c it upscale step by step to the original size, plus shotcircut connections

1 FPN(Feature Pyramid Network)

This is from Meta by YOLO inventor and Kaiming The network is similar to Unet but addiing information instead of concatenations This is applied to video as well, shows better results than ViT

2 RetinaNet

Retina is one-stage detection and is very fast, and it’s widely used in face detections. Retina is based on FPN and add both class and box detections

The loss is also the combinatino of two losses, class loss and regression loss(box loss)