You are on page 1of 16

Using Viola-Jones algorithm

with two-point features


for searching for man-made objects in the grass

Kornilov F.A., Kostousov K.V., Perevalov D.S.

Institute of Mathematics and Mechanics,


Ural Branch of RAS, TZSU-2011
http://tvcs2011.technicalvision.ru
Ekaterinburg
1. Problem
Required to find the images of objects of interest – “camomile-like” objects.
Objects of interest are partially occluded, and are located on a complex
background - the grass.

Object of interest Object of interest


in the grass
2. Available methods
1. A special algorithm
2. Neural networks
3. Geometric matching

4. Viola-Jones algorithm
5. LBP (Local Binary Pattern)
6. HOG (Histogram of gradients)

7. Generalized Hough transform


8. Bayesian model (Markov random field)
2. Ways to address
1. A special algorithm (???)
2. Neural networks (???)
3. Geometric comparison ("as separate circuits)

4. Viola-Jones algorithm (? How it works with occlutions)


5. LBP
6. HOG

7. Generalized Hough transform


8. Bayesian model (? "Slow")
3. Viola-Jones algorithm
Today it is one of the most often used algorithms to quickly find the front faces.

Algorithm is training on the base of labeled examples.


At the core of work is a mix of Haar-like features used in cascades of classifiers,
constructed by boosting.

The algorithm is "universal" – potentially it can be trained to recognize not only


faces, but any other class of objects which have “almost” fixed shape.
4. Viola-Jones algorithm
in the case of partial obstruction
In Barczak, Andre L.C. «Evaluation of a Boosted Cascade of Haar-Like
Features in the Presence of Partial Occlusions and Shadows for Real
Time Face Detection» (2004) shown that Viola-Jones algorithm works
good with occluded faces, if training was made on a partially occluded
examples.
5. "Camomile" - is not a face
The outer contour of the face and its internal features (eyes, nose, mouth) are
almost convex.

“Camomile" – has a non-convex outer contour, and has no significant internal


features.

So it makes sense to investigate the issue of search for “camomiles" some


other non-Haar features.
6. Two-point features
Consider a set of features, which consists of the difference in
brightness of two pixels (x1,y1) and (x2,y2)
for all possible pairs of pixels:

feature(x1,y1,x2,y2 ) = image(x1,y1) - image(x2,y2).

This set of features are significantly nonlocal.

This feature set is already used in several other works.


For example, in work on the recognition of gender with the faces:
Baluja, S., Rowley, H.A., Boosting sex identification performance.
Internat. J. Comput. Vision 71 (1), 111-119, 2007.
7. Which feature set is better?

or

Haar features Two-point features


7. Which feature set is better?
Training on 600 images
(«+»: 16x16 pixels, a small rotation of -2 .. 2 degrees;
«-»: 128x128 pixels).
Testing on 2000 of synthesized images (300x300 pixels, a several passes with
the rotation of images, by 5 degrees).

Synthesis of images: a layer of grass, an Examples of positive examples for


object circles the grass and structural training
noise (twigs and leaves), Gaussian (Blue channel, 16x16 pixels).
noise.
7. Which feature set is better?

Correct False alarm


detection
Haar features 84.2% 4.9%

Two-point 85.7% 7.3%


features

Output
Performance of both sets of features are similar,
just the two-point set is more sensitive.
8. Examples

Haar feature set Two-point set of features


8. Examples

Haar feature set Two-point set of features


8. Examples

Haar feature set Two-point set of features


8. Further research
1. What is performance of two-point features on frontal
faces?

2. How 3-and 4-point features will work?


Thank you for your attention
perevalovds @ gmail.com

You might also like