A summary of Rich feature hierarchies for accurate object detection and semantic segmentation

Ross Girshick et al. CVPR, 2014 DOI

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

Table of Contents

• A summary of Rich feature hierarchies for accurate object detection and semantic segmentation
  • 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 achieve SOTA object detection by using a combination of computer vision and deep learning techniques. This was chosen because object detection on the PASCAL VOC 2012 dataset results have stagnated for the past two years prior to publication.

Motivation

Why should we care about this paper?

We should care about this paper because the author’s propose an efficient method for performing object detection while also being more accurate than previous results. Furthermore, they introduce the usage of fine-tuning a larger CV model on a domain specific task in order to achieve better performance.

Category

What type of paper is this work?

This paper is both an algorithms paper and a CV model review paper.

Context

What other types of papers is the work related to?

This paper is related to works that talk about SOTA models, as well as papers that talk about CV models for object detection.

Contributions

What are the author’s main contributions?

Their main contributions are a proposed algorithm for performing object detection using DL based region proposals prior to CNN analysis and the usage of fine-tuning a larger model on a domain specific task to improve accuracy/ performance on the task.

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 creating CNNs for the PASCAL VOC challenge, creating algorithms for feature extraction for the purposes of object detection, creating algorithms for region proposal (selective search was used for this paper), and creating ensemble based DL models.

Figures, Diagrams, Illustrations, and Graphs

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

All of the figures and tables are easy to read and are clear.

Clarity

Is the paper well written?

The paper is well written. It goes into great technical length about the attributes of the model as well as how the model operates via an ablation study.

Relevant Work

Mark relevant work for review

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

2. N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In CVPR, 2005.
3. B. Alexe, T. Deselaers, and V. Ferrari. Measuring the objectness of image windows. TPAMI, 2012.
4. J. Uijlings, K. van de Sande, T. Gevers, and A. Smeulders. Selective search for object recognition. IJCV, 2013.
5. I. Endres and D. Hoiem. Category independent object proposals. In ECCV, 2010.
6. J. Carreira and C. Sminchisescu. CPMC: Automatic object segmentation using constrained parametric min-cuts. TPAMI, 2012.
7. P. Arbel´aez, J. Pont-Tuset, J. Barron, F. Marques, and J. Malik. Multiscale combinatorial grouping. In CVPR, 2014.
8. D. Cires¸an, A. Giusti, L. Gambardella, and J. Schmidhuber. Mitosis detection in breast cancer histology images with deep neural networks. In MICCAI, 2013.

Methodology

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

The authors compared their model (R-CNN) with and without bounding boxes against the previous SOTA models. Additionally, they conducted an ablation study on their model to understand how it works internally.

Author Assumptions

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

The authors don’t seem to make any assumptions.

Correctness

Do the assumptions seem valid?

There were no assumptions that I could find.

Future Directions

My own proposed future directions for the work

I know that this work is the predecessor to many works that expand and improve upon the R-CNN algorithm presented here. Therefore, I’d like to read those papers.

Open Questions

What open questions do I have about the work?

What is the performance of other region proposal algorithms when used in the R-CNN architecture?

Author Feedback

What feedback would I give to the authors?

This is a solid paper. I appreciate the alibation study conducted to understand the model.

Summary

A summary of the paper

The paper Rich feature hierarchies for accurate object detection and semantic segmentation by Ross Girshick et al. [1] introduces and evaluates the Region Convolutional Neural Network (R-CNN) on variations of the PASCAL VOC dataset for both object classification and semantic segmentation. They compare this algorithm against previous SOTA models on that dataset as well. Additionally, they perform an alibation study on their R-CNN model to understand how it works.

R-CNN works by using an ensemble approach. First, a DL model identifies regions of interest of an image. These regions can number up into the tens of thousands. These regions are then warped to fit a 224 x 224 pixel image. From this new image, it is passed into a traditional CNN model for object detection or semantic segmentation. By performing this region analysis first, the authors reduce the amount of data that the CNN needs to process, and therefore increases performance. Furthermore, the author’s pruned their R-CNN model and found that 94% of parameters could be dropped post-training.

Summarization Technique

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

Citations