A summary of You Only Look Once: Unified, Real-Time Object Detection

Joseph Redmon et al. CVPR, 2016 DOI

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

Table of Contents

• A summary of You Only Look Once: Unified, Real-Time 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?

The authors wanted to create a very fast object detection network that handles object detection using both region proposals and class probability maps in one unified model.

Motivation

Why should we care about this paper?

Because this model was the fastest object detection model of its time with being able to infrence at 45 FPS or 155 FPS.

Category

What type of paper is this work?

This is an object detection computer vision paper

Context

What other types of papers is the work related to?

This paper is related to work in real time object detection.

Contributions

What are the author’s main contributions?

Their main contribution is a model architecture (implemented with VGG) that is very fast at performing object detection in real time (45 FPS) or super fast (155 FPS) at the cost of accuracy.

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 to develop real time object detection systems as well as region proposal based object detection models.

Figures, Diagrams, Illustrations, and Graphs

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

The figures and tables in this paper are clear and easy to understand.

Clarity

Is the paper well written?

This paper is well written.

Relevant Work

Mark relevant work for review

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

Methodology

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

The authors tested variations of their model on the PASCAL VOC dataset against other SOTA models and measured the MAP percentage of the results. Furthermore, they measured the real time object detection performance of their model variations as well.

Author Assumptions

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

YOLO trades accuracy for speed which the author’s argue in their Introduction is applicable to the domain of self driving cars.

Correctness

Do the assumptions seem valid?

While infrence speed is a necessity within that domain, I would argue that accuracy is more important, as I would want a system that could detect a stop sign in front of the car accurately but slower, than quickly but innacurately.

Future Directions

My own proposed future directions for the work

This work has undergone many revisions since its initial publication. I’d like to review the enhancements that were made to the model since this initial publication and see what was changed and understand why.

Open Questions

What open questions do I have about the work?

Is this model bottlenecked by the number of classes that it has to look at and/or understand?

Author Feedback

What feedback would I give to the authors?

This was a really good paper. I would encourage further work in this field and specifically to test this model out on low powered devices.

Summary

A summary of the paper

The paper You Only Look Once: Unified, Real-Time Object Detection by Joseph Redmon et al. [1] describes an object detection strategy that aims to outperform previous methods in terms of infrence speed measured in FPS by creating a unified model that can perform region proposals and class probability mapping to a source image in parallel. This model is called YOLO.

The authors of this paepr were able to accomplis this by using the following technique:

1. Divide the source image into many sub-sections.
2. In parallel, compute the regions of interest as well as the class probability mapping for each sub-section.
3. Predict the class label and bounding boxes at the final layer by analyzing the overlap between the class probability mapping and regions of interest.

Their methodology was implemented using the VGG architecture + extra layers to accomodate for the parallel nature of the design.

Summarization Technique

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

Citations