[Paper note] Feature Pyramid Networks for Object Detection

• paper

Intuition

• Multi-scale is important in traditional methods
• Current detection system use single shot CNN to save time and memory (Faster R-CNN)
• CNN is capable of representing higher-level semantics, but not all levels are semantically strong
• Single Shot Detector (SSD) is one of the first attempts at using ConvNet’s pyramidal feature, but they add new layers after high up layer, which may lose information in high-resolution feature map
• Main contribution: perform multi-scale in the network

Model

• Feature Pyramid Network (FPN) building block
  
  • 
  • The feature maps in the picture above are from the last layer of each stage in ConvNet
  • Nearest neighbor upsampling
  • Denotes final set of feature maps as {P2,P3,P4,P5}, corresponding to {C2,C3,C4,C5}
• FPN in Region Proposal Network (RPN)
  
  • Replacing single-scale feature map with FPN
  • Anchors with different aspect ratios: 1:2, 1:1, 2:1
  • 15 anchors over the pyramid
• FPN in Fast R-CNN

Experiment

• Evaluate on COCO minival set
• Surpass 2016 COCO winner
• Lateral and top-down connection is helpful