0%

(ICCV 2017) Focal Loss for Dense Object Detection

Posted on Edited on In , ,

Keyword [Focal Loss]

Lin T Y, Goyal P, Girshick R, et al. Focal loss for dense object detection[C]//Proceedings of the IEEE international conference on computer vision. 2017: 2980-2988.


1. Overview

1.1. Motivation

  • two-stage approach contains sparse set of candidate object location
  • one-stage approach suffered from foreground-background class imbalance

In this paper, it proposed focal loss function to deal with class imbalance of one-stage approach (foucs on hard example [large error]).

1.2. Comparison

  • Two-stage
    • Proposal mechanism (1-2k proposals)
    • Biased minibatchsampling (foreground:background=1:3)
  • One-stage
    • 100k proposals (densely)
    • Most are easy negative which contribute no useful learning signal
    • Easy negative overwhelm
  • Two-Stage
    • R-CNN series
  • One-Stage
    • OverFeat
    • SSD
    • YOLO
  • Class Imbalance
    • Hard Negative Mining. completely discard easy example
    • Sampling Schemes
  • Robust Estimation
    • Huber Loss. reduce the contribution of outliers by down-weighting
    • Focal Loss. Down-weighting inliers

1.4. Focal Loss



  • α. balance factor (inverse class frequency or hyperparameter setted by cross validation) improved accuracy over non-α form
  • γ. focusing parameter, equal to 2 better for experiments

When summed over a large number of easy examples, small loss values can overwhelm the rare class.



1.4.1. Initialization

Introduce prior concept for the model estimation probability in the beginning. set final layer’s



means that p = π when get through sigmoid function. (π = 0.01 in experiments)

1.5. RainNet Detector




2. Experiments

2.1. Effect



2.2. Ablation Study



2.3. Experiments