---

Stacked Specialist Regressors (SSR)

Student entry for the COCO Keypoints Challenge 2016.

Team : C. Sutton, R. Canyasse, G. Demonet

Abstract : TODO

Table of contents :

• Introduction
• Installation
• Quick start
• The idea behind
• Some potential improvements
• References

Introduction

TODO

Installation

TODO

Quick start

TODO

The idea behind

TODO

Some potential improvements

TODO

References

TODO

---

---

---

GitHub release to-do list :

• Prepare project build structure
• Clean and adapt code
• Write new necessary code for sharing (experiment reproducibility, results...)
• Document the reproduction and skeleton of the code
• Complete README with background and potential leads forward

---

1. Project build structure

Several components :

• Data acquisition and handling

  • Links to MS COCO dataset and challenge
  • Script to automatically populate the data folders if empty
  • Image batches generators for Keras models

• Person detection task : Faster-RCNN (thanks to github.com/rbgirshick)

  • Model description and links
  • Script to generate bounding boxes for test images

• Keypoints regression task : SSR model

  • Model definition : keras implementation
  • Training script
  • Prediction script

• Evaluation and results

  • Model output handling + "aggregation" method
  • Visualizations
  • Graphs and performance/time/space reports

The repo structure is then :

• data/

  • coco/
  • scripts/
  • README.md

• frcnn/

  • ???/
  • README.md

• ssr/

  • model/
  • README.md

• eval/

  • scripts/
  • README.md

• README.md