How to convert the accuracy of the measurement results of the SIFT algorithm

The SIFT algorithm is broken down into four steps:

2.Keypoint Positioning: At each candidate location, a finely fitted model is used to determine the location and scale. Key points are selected based on how stable they are.

3.Direction Determination: Assign one or more directions to each key location based on the local gradient direction of the image. All subsequent operations on the image data are transformed relative to the direction, scale, and position of the keys, providing immutability to these transformations.

4.Key Description: Measure the gradient of the image locality on a selected scale, within the neighborhood around each key. These gradients are transformed into representations that allow for larger local shape distortions and lighting variations.

For the scale space, the space is formed by the convolution of the i(x,y) expression of the original image and the Gaussian function g(x,y,delta).

This scale should be understood as a scale factor delta, which is used to control the scale of the image in the x and y directions, and it forms a space with a continuous function expression of x and y parameters l(x, y, delta), which is called the scale space, which is formed by the convolution of the i(x, y) expression of the original image and the Gaussian function g(x, y, delta). , just like the patent infringement compensation case of *** chip back then.

This kind of question is really difficult to play, what is the p here? How much is r? 6p=7r, which is equal to seven, and I am a junior high school student. I don't know this question, you, do you know the meaning of this question? How can you solve it?

Did you solve it in the end? Find the ** of calculating the angle! Please, please.

Theoretically, SIFT is a similarity invariant, i.e., the change in image scale and rotation are invariants. However, due to the special processing of many details when constructing SIFT features, SIFT has strong adaptability to the complex deformation and illumination changes of the image, and the calculation speed is relatively fast and the positioning accuracy is relatively high. Such as:

Compared with the traditional detection method based on log operator, the operation speed is greatly accelerated.

The precise positioning of key points not only improves the accuracy, but also greatly improves the stability of the key points;

In the construction of descriptors, the statistical characteristics of the sub-region instead of a single pixel are used as the research object, which improves the adaptability to the local deformation of the image.

For the 16*16 key point neighborhood and the 4*4 sub-region, the gradient amplitude is weighted like the Gaussian function, which strengthens the central region and dilutes the influence of the edge region, so as to improve the adaptability of the algorithm to geometric deformation.

This method is not only invariant to the general linear lighting model, but also adaptable to complex lighting changes. For more information on this section, see the article "Distinctive image features from scale-invariant keypoints".

Characteristics of SIFT algorithm: SIFT feature is a local feature of the image, which maintains invariance to rotation, scale scaling, and brightness changes, and also maintains a certain degree of stability to angle changes, affine transformations, and noise. It has good distinctiveness and rich information, which is suitable for fast and accurate matching in a massive feature database. Multiplicity, even a few objects can generate a large number of SIFT eigenvectors; High speed, optimized SIFT matching algorithm can even meet real-time requirements; Extensibility, which can be easily combined with other forms of eigenvectors.

The main process is as follows: (1) randomly select a ransac sample from the sample set, i.e., 4 matching point pairs, (2) calculate the transformation matrix m according to these 4 matching point pairs, (3) calculate the consensus of the congregate set satisfying the current transformation matrix according to the sample set, the transformation matrix m, and the error metric function, and return the number of elements in the congruence set (4) according to the elements in the current congruent set.

In the first two articles I wrote about the SIFT algorithm, I have already had a preliminary introduction to the SIFT algorithm: 9. SIFT algorithm for image feature extraction and matching, and then in the compilation and implementation of SIFT algorithm (IX) (continued), I also briefly recorded how to use OpenCV, GSL and other libraries to compile and run SIFT programs.

However, according to a friend, whether it is possible to implement the SIFT algorithm in C language, and at the same time, try not to use things such as OpenCV, GSL and other third-party libraries. Moreover, the SIFT library maintained by Rob Hess is not easy to understand, and some people don't understand what is going on.

So this article will teach you how to use C language to implement the SIFT algorithm step by step, and at the same time, you will be able to really understand what the SIFT algorithm is all about.

The information of SIFT feature points has coordinates, directions, and gradient metrics, and you can create a feature point class to encapsulate all those information

For specific suggestions, please refer to the predecessors' programs, and it is a bit difficult to make it up by yourself!

This MATLAB is an empty shell, and its core is calling. Q1 (for Windows) This is the format requirement for PGM files. You can check Q2 by yourself!

Indicates that the commands that are executed are executed in the system shell. For Windows, it's the Command Prompt.

Q3 That's the parameter that the executable requires.

Q4 That's what the executable needs.

This means that this is not a matlab**, which means that it is useless to look at this, I also looked at this**structure, I recommend you to take a look at opencv**, someone on the Internet has sent a detailed explanation.

The result of the Gaussian difference is dog (difference of gaussian), which is an approximation of log (laplacian of gaussian).

The log image is by far the best scale transformation and is the most accurate.

However, because it is very difficult to calculate the log image, the author Lowe made an approximation of the log image, and he found that if he used dog, it was only a constant different from the original log, so that the position of the maximum and minimum value was unchanged.

And our goal is to find this maximum and minimum value and then find the location of this point of interest.

The point is that dog is very easy to calculate, as long as the two ** are subtracted.

That's why we use the difference of the Gaussian diagram!

Take a good look at ** is the key! **It's very detailed, I've read it in two days