Faster R-CNN 논문 리뷰

2025.01.04 - [DeepLearning/Object Detection 논문] - Fast R-CNN 논문 리뷰

Fast R-CNN 논문 리뷰

 

지난 Fast R-CNN을 소개했고 이를 더욱 성능을 극대화 시키도록 제작한 Faster R-CNN을 소개하겠습니다.

 

Faster R-CNN

 

• Bring the Region Proposal Network(RPN) into the framwork
  • No longer off-the-shelf proposal network needed.(기존에 사용했던 proposal model이 아닌 neural network를 사용) 

    

• Stage 1: Training the RPN(Region Proposal Network)
  • For each position in conv feature map, suppose a set of candidate bounding box(called Anchor).
    • Usually, 3 ratio (1:1, 1:2, 2:1) and 3 sizes, total 9 anchors used.(비율 3가지(파란색 박스), 크기 3가지(파,빨,녹))

      

      • ⁕ Anchor Examples
        • IoU: 예측한 boundig box와 GT의 교집합/합집합

          

  • • If the anchor sufficiently overlaps with a GT bounding box, it is considered as a positive.
    • If the anchor does not overlap with any GT bounding box, it is considered as a negative.
    • RPN is trained to classify positive/negative for each anchor and to regress the bounding box if it is a positive.
    • 중요!! Loss function for training RPN(예측proposal과 GT로 Anchor를 이용한 최종 proposal를 찾아냄)

      

      
      

• Stage 2: Same as Fast R-CNN
  • Using the RPN trained in Satage 1, perform Rol Pooling, classification, and regression.

출처:

https://www.youtube.com/watch?v=W6EVlzVP0TM&ab_channel=JoonseokLee

 

논문:

https://arxiv.org/abs/1506.01497