A summary of Focal Loss for Dense Object Detection

Tsung-yi Lin et al. arXiv, 2018 DOI

For the summary of the paper, go to the Summary section of this article.

Table of Contents

• A summary of Focal Loss for Dense Object Detection
  • Table of Contents
  • First Pass
    • Problem
    • Motivation
    • Category
    • Context
    • Contributions
  • Second Pass
    • Background Work
    • Figures, Diagrams, Illustrations, and Graphs
    • Clarity
    • Relevant Work
    • Methodology
    • Author Assumptions
      • Correctness
    • Future Directions
    • Open Questions
    • Author Feedback
  • Summary
  • Summarization Technique
  • Citations

First Pass

Read the title, abstract, introduction, section and sub-section headings, and conclusion

Problem

What is the problem addressed in the paper?

This paper aims to address the problem that one stage object detectors (i.e., YOLO, SSD) face when trying to match the performance of SOTA two stage object detectors which is class imbalance.

Motivation

Why should we care about this paper?

Because it introduces a new loss function that addresses the issue of class imbalance when training dense, one stage object detectors. Additionally, the authors released an example model implementing this loss known as Detectron/ RetinaNet.

Category

What type of paper is this work?

This is an algorithms and CV object detection paper.

Context

What other types of papers is the work related to?

This paper is related to papers demonstrating or working on one stage object detection models.

Contributions

What are the author’s main contributions?

The author’s main contribution is a new loss function aimed at training one stage object detection models that reduces the problem of class imbalance between identifying objects in the foreground and background. Furthermore, the authors have released an example model that was trained on this loss function known as Detectron/RetinaNet.

Second Pass

A proper read through of the paper is required to answer this

Background Work

What has been done prior to this paper?

Work has been done in developing classic object detectors, one and two stage detectors, reducing class imbalance, and robust estimation techniques.

Figures, Diagrams, Illustrations, and Graphs

Are the axes properly labeled? Are results shown with error bars, so that conclusions are statistically significant?

The figures are clear and understandable

Clarity

Is the paper well written?

This paper is well written and is dense with technical information.

Relevant Work

Mark relevant work for review

The following relevant work can be found in the Citations section of this article.

2. P. Doll ́ ar, Z. Tu, P. Perona, and S. Belongie. Integral channel features. In BMVC, 2009.
3. P. F. Felzenszwalb, R. B. Girshick, and D. McAllester. Cascade object detection with deformable part models. In CVPR, 2010.
4. T. Hastie, R. Tibshirani, and J. Friedman. The elements of statistical learning. Springer series in statistics Springer, Berlin, 2008.
5. W. Liu, D. Anguelov, D. Erhan, C. Szegedy, and S. Reed. SSD: Single shot multibox detector. In ECCV, 2016.

Methodology

What methodology did the author’s use to validate their contributions?

They compared their RetinaNet model against other SOTA object detectors on the COCO dataset. Additionally, they compare the performance of models trained using their Focal Loss and the Online Hard Example Mining (OHEM) technique.

Author Assumptions

What assumptions does the author(s) make? Are they justified assumptions?

Their assumption is that one stage object detectors are the future.

Correctness

Do the assumptions seem valid?

While having more options as to what type of object detector to choose from (one or two stage), it is important to keep in mind that inference speed, accuracy, recall, other metrics, and domain need all play an important role in what model is selected for a particular task.

Future Directions

My own proposed future directions for the work

I’d like to implement Focal Loss in both a traditional YOLO network and a YOLO network following the MobileNet architecture.

Open Questions

What open questions do I have about the work?

Why was the COCO dataset chosen and not the ImageNet or Pascal VOC dataset for training?

Author Feedback

What feedback would I give to the authors?

This is a good paper. I would say that the size of the network is certainly larger than previous one stage object detectors such as YOLO. Could it be possible to reduce the size of the network to be comparable to these smaller networks while maintaining the accuracy or achieving a better accuracy?

Summary

A summary of the paper

The paper Focal Loss for Dense Object Detection by Tsung-Yi Lin et al. [1] describes a new loss function aimed at improving the performance of one shot object detection models that rely on region proposals. The problem that one shot object detection models face compared against traditional two stage models that utilize region proposals and object detection is that of class imbalance. Class imbalance is simply that the region proposal network detects too many regions where an object might be. This affects the performance of the object detection component of the model as it might infer that an object is in a location that it isn’t.

To reduce this error, the authors of this paper propose the Focal Loss function, a loss function aimed at reducing class imbalance. The function is FL(pt) = −(1 − pt)^γ log(pt) where γ >= 0 They then trained a model (RetinaNet) with this loss function on the COCO dataset and found that it performed better than other on stage methods with respect to average precision.

Summarization Technique

This paper was summarized using a modified technique proposed by S. Keshav in his work How to Read a Paper [0].

Citations