1. Introduction

• Feature pyramid in FPN
  • Before feature fusion
  • Top down feature fusion
  • After feature fusion

Design defects FPN

1. Semantic gaps between features at different levels

• 1x1 convolution does not consider semantic gaps between feature maps
• This degrades the power of multi-scale feature representation due to inconsistent semantic information

2. Information loss of the highest-level feature map

• Low level features improved with stronger semantic information, but highest pyramid level loses information due to reduced channels

3. Heuristical assignment strategy of RoIs

• Stage mapping is done based on scales of proposals heuristically
• Other level features may be beneficial
• 비슷한 scale을 가진 RoI들이 각각 다른 feature map에 대응 될 수 있음 → discrete, hard label

2. Related Work

Deep Supervision