/photogrammtry_midterm

Primary LanguageMATLAB

Estimating Dichromatic Reflection Model From Images

(A)Brief Description of Algorithm

As shown below, the reflection of light can be considered to be a combination of diffuse and specular reflections: In Dichromatic Reflection Model, incident light reflected from an object is written as a linear combination of diffuse and specular reflections: There are many ways to compute them, I choose Gauss-Seidel iterations to get an estimation of the reflection. In my program, I mainly utilized the VLFeat to get the superpixel segmentation pictures of my data. After I got segementation pictures, I utilized Gauss-Seidel iterations to get a relatively accurate result of color estimation.

The formulu which I used was shown as following:

(B)About Program

  • My program mainly has two part: superpixel segementation & Gauss-Seidel iterations
  • In my main function, I first included the path of VLFeat which I utilized to get superpixel and do segmentation
  • Then a function named GS_iter was called. Its main idea is to conducted Gauss-Seidel iterations on the segmentation results to get relatively accurate value of md cd ms cs (which 'd' means 'diffuse' and 's' means 'specular'). During this process two function were also called: cal_m & cal_c which were used to calculate the color value and its coefficient.
  • Finally, in order to get the final result of true color picture, I used each channel of cd multiply with md

(C)Result

"Duck"

the picture is listed as following order:

original + final_result -> original_superpixel_segmentation + changed_superpixel_segmentation -> md + ms

We can see from the result and original_superpixel_segmentation that there still exists some segementation mistakes (The head of duck should be a complete part, but in original_superpixel_segmentation the color was divided into different parts). Besides, the eye's color is also wrong in the result. So I changed the iteration times from 3 to 10, then I got a relatively worse result: Compared with result shown above, we can see the eye's color even become as same as its head.

"RPAC basketball playground"

Although there exists some shortcomings in my outcome, I also get some good results: As the result shows, the color of floor after optimization is better compared with original pictures.

                                                                                                kong.325