Why do we need a neck?

• In order to detect a variety of sizes of objects
• High level features have strong semantic information but less raw pixel data and low level features have strong pixel data and less semantic information.
• A neck allows high and low level features to exchange information

1. Neck

1-1. Feature Pyramid Network

• high-level 에서 low-level로 semantic 정보 전달
• top-down pathway 추가

Lateral connections

Pipeline

• backbone : ResNet
• stage mapping을 위한 함수 → feature 가 어디서 왔는지 계산

1-2. Path Aggregation Network (PANet)

• Adaptive Feature Pooling