mitemitreski / javasurf Goto Github PK

On any image I apply the library to, features appear in the lower right corner 
in a stationnary position

What steps will reproduce the problem?
1. take a pure white 512*512 png image
2.  threshold = 2000;
    balanceValue = 0.9;
    octaves = 8;
3. Run the algorithm/test suite against this image

The lower right corner contains irrelevant features, they are present even on 
non white images. (And if there is pattern, they don't move relative to the 
image even if the pattern rotates). This makes it very annoying when using the 
library to match and overlay 2 similar images.

The number of those detections increase with the number of octaves.




What version of the product are you using? On what operating system?


Please provide any additional information below.

First notice that when computing IntegralSquare, point A is excluded and point 
D is included (you got sum(pixel) for x>xA && x<=xD && y>yA && y<=yD)

This is fine because point coordinates are computed using (int)(x-scale) for 
point A and (int)(x+scale) for point D resulting in truncation (except if scale 
if an integer value).

Looking at HAARConvolution.makeConvolutionDx for example, we can see that the 
width of the positive wavelet is 1 pixel greater than the width of the negative 
one. This is because the positive wavelet include the pixels values for the 
center of the filter. This is not good as the filter there is 0.

Suggest to modify code as follow:


  return (  integralImage.getIntegralSquare(x1, y1, x2-1, y3) 
          - integralImage.getIntegralSquare(x2, y1, x3, y3));

Same comment for other methods in HARRConvolution and GAUSSConvolution.