RetinaNet

Focal Loss for Dense Object Detection

Introduction

本文提出了一种 Focal Loss
这种损失函数更关注 hard, misclassifed examples

为了验证这种 Focal Loss 的效果
我们设计了一个1-stage的RetinaNet
可以看出其精度已经超过了FPN Faster R-CNN
同时拥有1-stage的速度

Related Work

Classic Object Detectors
LeCun的手写数字识别是最早的分类工作之一
Viola和Jones用boosted目标检测器做人脸检测
HOG和integral channel features提升了行人检测效果

Two-stage Detectors
R-CNN
Faster R-CNN

One-stage Detectors
OverFeat
SSD
YOLO

Class Imbalance
对于1-stage的算法
These detectors evaluate 10⁴ -10⁵ candidate locations per image but only a few locations contain objects.
(1) training is inefficient as most locations are easy negatives that contribute no useful learning signal
(2) enmasse, the easy negatives can overwhelm training and lead to degenerate models.
我们的Focal Loss对这个问题有着很好的改善

Robust Estimation
与Huber Loss不同
the focal loss performs the opposite role of a robust loss: it focuses training on a sparse set of hard examples.

Focal Loss

可以看出 pt 越接近1 分类越准确，损失就越小
这样大量的负样本总体的损失就大大减小，防止正类损失被淹没

we found γ = 2 to work best in our experiments

Class Imbalance and Model Initialization
模型初始化解决类别平衡问题

RetinaNet Detector

前面用到了FPN
后面是两个子网络分别用于分类和回归
同时也用到了RPN中的平移不变性anchor

回归网络预测目标和R-CNN一致
用ResNet在ImageNet上的训练参数初始化网络

Experiments