Introduction

The McGill Object Detection Suite (MODS) is a software package for creating test sets suitable for evaluating object detection systems.  These test sets are created by superimposing objects from existing publicly available object databases (currently, COIL-100 and SOIL-47) onto heterogeneous backgrounds.  Test sets focusing on pose, in-plane rotation, scale, illumination, occlusion, or noise can all be created with the MODS.  This software package is being made publicly available to aid the computer vision community by providing standard test sets which will allow object detection systems to be systematically compared and characterized.
 

Sample Images from Test Sets

Sample images at half their actual size (250 x 250 pixels instead of 500 x 500 pixels) are shown here for each type of test set (pose, in-plane rotation, scale, illumination, occlusion, and noise) .  The MODS can also produce test sets combining these factors (i.e., images with objects containing both out-of-plane and in-plane rotation).  Objects are taken from both the COIL-100 and SOIL-47 databases.
 

Pose (Out-of-Plane) Rotation

 
Objects 9 and 12 from the COIl-100 database with an out-of-plane rotation of 10º and 60º, respectively, between successive object instances.
 

 
Objects 3 and 8 from the SOIL-47 database with an out-of-plane rotation of 18º between successive object instances.

In-Plane Rotation

  
Objects 18 and 92 from the COIl-100 database with an in-plane rotation of 60º between successive object instances.

 
Objects 3 and 8 from the SOIl-47 database with an in-plane rotation of 30º and 45º, respectively, between successive object instances.

Scale

 
Object 64 from the COIL-100 database shown at a scale of s = 3.6 (height ~= 460 pixels) and s = 0.28125 (height = 36 pixels).

 
Object 1 from the SOIL-47 database shown at a scale of s = 4 and s = 0.2.

Occlusion

 
Object 2 from the COIL-100 database artificially occluded by approximately 25% and 50%.

 
Object 7 from the SOIL-47 database artificially occluded by approximately 25% and 75%.

Illumination (Alternate Illumination Condition)

   
Object 16 from the SOIL-47 database under two different illumination conditions.

Illumination (Contrast Change)

   
Object 1 from the COIL-100 database with a contrast change of 0.5, 1.0 (nominal lighting condition), and 1.25, respectively.

Illumination (Brightness Change)

 
Object 1 from the COIL-100 database with a brightness change of -0.3 and 0.3, respectively.

Illumination (Spotlight)

 
Object 1 from the COIL-100 database with a spotlight level of 1.5 and 2.0, respectively.

Noise

   
Object 17 from the COIL-100 database with no noise, Gaussian noise of N(μ=0, σ2=0.01), and Uniform noise of U(-0.2, 0.2).  Each colour channel of a pixel has a value between 0 and 1.

 
Object 6 from the SOIL-47 database with no noise and Salt and Pepper noise where each pixel is set to black or white with probability p=1%.


Combined


Object 20 from the COIL-100 database with object instances rotated both out-of-plane and in-plane.  In addition, the object is occluded by 25% and Salt and Pepper noise (p=1%) has been added to the image.  This is a challenging test image for object detection systems.
 

Downloads

Revision History

Version Modifications
1.1
  • Added ability to artificially illuminate object instances (contrast, brightness, and spotlight features).

  • Added ability to modify which background images are used

1.0
  • Initial Release

 

Useful Links

Acknowledgements

The authors would like to thank Gwen A. Williams for her grammatical insights and aid in producing the COIL-100 and SOIL-47 masks.  The authors gratefully acknowledge funding provided by the Natural Science and Engineering Research Council of Canada.    


Use Policy

Permission is hereby granted to use the McGill Object Detection Suite for academic purposes only, provided that the suite is referenced in publications related to its use as follows:

Donovan H. Parks and Martin D. Levine, "The McGill Object Detection Suite", Third Canadian Conference on Computer and Robot Vision, CRV2006, June 7-9, 2006, Québec, Canada, 2006.


Contact Info

This software package is still in development and any suggestions regarding the McGill Object Detection Suite are appreciated.  Comments and suggestions should be directed to: Donovan Parks [donovan.parks (at) gmail.com].


Last modified: April 19, 2007 by D. H. Parks
Created: February 1, 2006 by D. H. Parks