Skip to content

Latest commit

 

History

History
28 lines (21 loc) · 1.46 KB

efficientdet.md

File metadata and controls

28 lines (21 loc) · 1.46 KB

EfficientDet: Scalable and Efficient Object Detection

September 2021

tl;dr: BiFPN and multidimensional scaling of object detection.

Overall impression

This paper follows up on the work of EfficientNet. The FPN neck essentially is a multi-scale feature fusion that aims to find a transformation that can effectively aggregate different features and output a list of new features.

Key ideas

  • BiFPN (bidirectional FPN) (<-- PANet <-- FPN)
    • PANet to introduce bottom up pathway again.
    • Remove nodes from PANet that has only has one input edge.
    • Add skip connection from original input to the output node if they are at the same level
    • Repeat blocks of the above BiFPN block.
  • Weighted feature fusion
    • Baseline is to resize and sum up. Each feature may have different weight contribution (feature level attention).
    • Softmax works, but a linear weighting normalization may work as well.
  • Multidimensional/compound scaling up is more effective than single dimension scaling. Resolution, depth and width.

Technical details

  • NAS-FPN has repeated irregular blocks.
  • Simply repeating FNP blocks will not lead to much benefit. Repeating PANet blocks will be better, and repeated BiFPN yields similar results but with much less computation.
  • This still needs object assignemtns, like RetinaNet.

Notes