BoofCV: Real Time Computer Vision in Java

Introduction 

BoofCV is a new real time computer vision library written in Java.  Written from scratch for ease of use and high performance, it provides a range of functionality from low level image processing, wavelet denoising, to higher level 3D geometric vision.  Released under an Apache license for both academic and commercial use.  BoofCV's speed has been demonstrated in a couple of comparative studies against other popular computer vision libraries (link).

This article will demonstrate how to detect image features and associate two images together.  Image association is a vital component in creating image mosaics, image stabilization, visual odometry, 3D structure estimation, and many other applications.  BoofCV's website contains several other tutorials, applets, and example code. 

Since this article was written BoofCV is at version 0.7 as of April 2012.  With many new features and performance improvements!

Website: http://boofcv.org
Version: Alpha v0.2
Date: December, 1 2011
Author: Peter Abeles  

Binary Image Processing	Image Registration and Model Fitting	Interest Point Detecting

Background  

It is assumed that the reader is familiar with basic concepts in computer vision as well as Java development.  BoofCV is a very new library, with its first public release in early November 2011.  It is undergoing a rapid phase of development.  If you like BoofCV please let other people know about it by sharing its webpage with your friends! 

Browsing through the code below you might notice that BoofCV makes extensive use of Java generics.  Generics allows BoofCV to maintain strong typing and provide easy to use highly abstracted data types.  It is recommend that you are familiar with generics before trying to use BoofCV.

Image Registration Example  

BoofCV provides several different ways to register images.  Most of them fall under the category of interest points.  In this context, an interest point is a feature inside the image which can be easily and repeadily recognized between multiple images of the same scene from different points of view.    If Java is set up in your browser, then you can see feature association in action by taking a look at this applet:

 Feature Association:  http://boofcv.org/index.php?title=Applet_Associate_Points

In the example, below the two images are registered to each other in several steps:

1. Detect interest points
2. Describe interest points 
3. Associate image features  

In the block of code below the class is defined and several classes are passed in.  These classes are abstract interfaces which allow several algorithms to be swapped in for each other.  New ones can be easily added in the future.  While not shown in this example, the un abstracted code is also easy to work with when high performance is required over easy of development.

public class ExampleAssociatePoints<T> {
 
	// algorithm used to detect interest points
	InterestPointDetector<t> detector;
	// algorithm used to describe each interest point based on local pixels
	DescribeRegionPoint<t> describe;
	// Associated descriptions together by minimizing an error metric
	GeneralAssociation<tupledesc_f64> associate;
 
	Class<t> imageType;
 
	public ExampleAssociatePoints(InterestPointDetector<T> detector,
				DescribeRegionPoint<T> describe,
				GeneralAssociation<TupleDesc_F64> associate,
				Class<t> imageType) {
		this.detector = detector;
		this.describe = describe;
		this.associate = associate;
		this.imageType = imageType;
	}

	
	 /**
	 * Detect and associate point features in the two images.  Display the results.
	 */
	public void associate( BufferedImage imageA , BufferedImage imageB )
	{
		T inputA = ConvertBufferedImage.convertFrom(imageA, null, imageType);
		T inputB = ConvertBufferedImage.convertFrom(imageB, null, imageType);
 
		// stores the location of detected interest points
		List<Point2D_F64> pointsA = new ArrayList<Point2D_F64>();
		List<Point2D_F64> pointsB = new ArrayList<Point2D_F64>();
 
		// stores the description of detected interest points
		FastQueue<TupleDesc_F64> descA = new TupleDescQueue(describe.getDescriptionLength(),true);
		FastQueue<TupleDesc_F64> descB = new TupleDescQueue(describe.getDescriptionLength(),true);
 
		// describe each image using interest points
		describeImage(inputA,pointsA,descA);
		describeImage(inputB,pointsB,descB);
 
		// Associate features between the two images
		associate.associate(descA,descB);
 
		// display the results
		AssociationPanel panel = new AssociationPanel(20);
		panel.setAssociation(pointsA,pointsB,associate.getMatches());
		panel.setImages(imageA,imageB);
 
		ShowImages.showWindow(panel,"Associated Features");
	}

 
	/**
	 * Detects features inside the two images and computes descriptions at those points.
	 */
	private void describeImage(T input, List<point2D_F64> points, FastQueue<tupledesc_f64> descs )
	{
		detector.detect(input);
		describe.setImage(input);
 
		descs.reset();
		TupleDesc_F64 desc = descs.pop();
		for( int i = 0; i < detector.getNumberOfFeatures(); i++ ) {
			// get the feature location info
			Point2D_F64 p = detector.getLocation(i);
			double yaw = detector.getOrientation(i);
			double scale = detector.getScale(i);
 
			// extract the description and save the results into the provided description
			if( describe.process(p.x,p.y,yaw,scale,desc) != null ) {
				points.add(p.copy());
				desc = descs.pop();
			}
		}
		// remove the last element from the queue, which has not been used.
		descs.removeTail();
	}

  
	public static void main( String args[] ) {
 
		Class imageType = ImageFloat32.class;
 
		// select which algorithms to use
		InterestPointDetector detector = FactoryInterestPoint.fastHessian(1, 2, 400, 1, 9, 4, 4);
		DescribeRegionPoint describe = FactoryDescribeRegionPoint.surf(true, imageType);
		GeneralAssociation<TupleDesc_F64> associate = FactoryAssociation.greedy(new ScoreAssociateEuclideanSq(), 2, -1, true);
 
		// load and match images
		ExampleAssociatePoints app = new ExampleAssociatePoints(detector,describe,associate,imageType);
 
		BufferedImage imageA = UtilImageIO.loadImage("../evaluation/data/stitch/kayak_01.jpg");
		BufferedImage imageB = UtilImageIO.loadImage("../evaluation/data/stitch/kayak_03.jpg");
 
		app.associate(imageA,imageB);
	}
}

Image above shows pairs of detected and associated interest points inside two images at different orientations.  That's it for now!

This article, along with any associated source code and files, is licensed under The Apache License, Version 2.0