Learning-Deep-Learning

NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Detection

May 2019

tl;dr: Using NAS to search for best cross-scale connection in FPN.

Overall impression

First paper to use NAS to search optimal pyramid architecture for object detection. NAS-FPN finds scalable design which better benefits from repeating the modules. However it requires quite a lot of computational resource to search and not easily reproducible in small labs. Most manly engineered cross-scale connections are shallow (only connecting neighboring or the same scales) but NAS-FPN can connect non-neighboring scales as well.

NAS has many other usage, such as Auto Augment.

Key ideas

Technical details

Notes