task : dense object detection with localization score
problem : 기존의 localization score를 뽑는 work들은 1) 학습할 때 IoU score 브랜치를 따로 학습하다가 infer할 때 결합해서 train - infer 간의 괴리가 생겼고 2) localization quality가 positive에만 부과되어서 negative sample도 IoU score가 매우 높게 나오는 경우가 있었으며 3) bbox 분포에 대한 가정이 Dirac-Delta이거나 gaussian으로 너무 단순했다.
idea : 학습 때 category와 IoU score를 결합하여 smooth한 target을 주어 학습-추론 괴리를 없애고 bbox에 대한 분포도 학습하게 하여 분포의 강한 제약을 없애자
architecture : ResNet with FPN + ???
objective : 1) focal loss의 $(1-p_t)^\gamma$ 대신에 target 과의 거리 term인 $|y-\sigma|^\beta$를 곱해주고 2) 학습한 discrete 분포에 대한 값도 반영해줌 => Generalized Focal Loss
paper
TL;DR
Details
기존 방법의 문제점
Generalized Focal Loss의 주요 아이디어
focal loss![image](https://user-images.githubusercontent.com/46675408/202345361-561dfebd-9beb-4c7c-8e28-b3dd09ac426e.png)
quality focal loss![image](https://user-images.githubusercontent.com/46675408/202345533-4c98f024-c846-4684-ad9e-76f6ccad2638.png)
distribution focal loss![image](https://user-images.githubusercontent.com/46675408/202345867-9b049a8c-e2a1-42ad-93e9-fbc2ec85a576.png)
Result