Opencv learning notes () -- Introduction to the new version 2.4.2 & comparison of features2d and freak and ORB feature descriptions

Opencv 2.4.2 Introduction

Opencv never slowed down its own footwork,In July 4, the latest version 2.4.2 was released again.. At the same time, two website regions are released.

Let's take a look at the updates and changes to the new version:

Now supports IOS OS! You can also use opencv to develop your iPhone.

Added a new feature description.AlgorithmFreakThis is also a new cvpr algorithm this year. It is also a 2-value feature. According to the principle of human retina, it is quite mysterious. It is said that the effect is better than orb and surf, I want to try it.

The GPU module is further optimized to greatly improve the performance of individual algorithms.

Fixed more than 50 bugs in 2.4.1.

There are so many introductions. I feel that the highlights are the website construction and the freak feature. Other things have to be explored.

Comparison between freak and ORB feature description sub-effects

Orb is the improvement of brief. Brief is too simple to introduce. If you are interested, read paper. After reading the orb paper, I felt that the improvements were mainly as follows:Using fast as the algorithm for Extracting Feature Points is faster, and the main direction of the feature points is added, thus having rotation immutability. Finally, I also thought about it. At that time, I thought we should be able to optimize brief, that is, ORB adopted the greedy and exhaustive method to obtain random points with low correlation, another improvement is that for random points, the brief method is to filter the original image and reduce the noise. Orb does not directly compare the pixels, instead, a small patch centered on the pixel is selected as the comparison object, improving the anti-noise capability..

I just read the freak and may not understand it,In my understanding, this algorithm is based on the distribution of human eyes and retina cells, which are dense in the middle and sparse around, so as to build many areas in the image. Of course, the closer the center area, the more intensive the sampling, the sampling in the surrounding area is sparse, and a set of two-value features are obtained by randomly comparing the pixels in each area. This algorithm also focuses on the scale and direction issues and provides corresponding solutions, based on the constant eye rotation when people see things, a cascade searcher is designed.All in all, I think this algorithm is also an improvement inspired by the two-value feature orb and brisk.

Next we will use opencv to compare it. It is worth noting that in 2.4.2,Freak provides pattern training.Code.

The comparison results are as follows: (ORB, freak)

The filtered feature points are as follows: (ORB, freak)

Of course, I am just a simple test here. In fact, it cannot fully reflect the actual performance,Here, freak uses surf to detect the feature points and ORB uses fast.. However, from the comparison results here,Orb is much better. Next let me compare the running time of the algorithm:

	Orb	Freak Detection	Freak Extraction
Debug	0.317 s	0.301 s	0.167 s
Release	0.094 s	0.162 s	0.085 s

Although the two algorithms cannot be directly compared, freak does not provide an algorithm for detecting the feature point location,I personally feel that if freak uses fast for detection, the speed should be faster than orb..

My Project Link: http://download.csdn.net/detail/yang_xian521/4421537