Generate MATLAB saliency masks for the DMSRA and MSSRA detectors for the complete process (color -> gray -> binary -> regions)

[elboyran]

From @elboyran on April 26, 2016 11:47

Same gray-image as for the binarization tests.

Copied from original issue: NLeSC/SalientDetector-python#30

[elboyran]

The gray image to be used will be the boat1 image. Cool car wasn't really gray.

[elboyran]

Parameters used for the DMSR: SE_size_factor = 0.02; Area_factor_very_large = 0.01; Area_factor_large = 0.001; lambda_factor = 3; num_levels = 255; offset = 80; otsu_only = false; conn = 8; weight_all = 0.33; weight_large = 0.33; weight_very_large = 0.33; saliency_type = [1 1 0 0];

Some computed parameters: lambda (for the binarization): 24 SE_size = 8 lambda (for DMSR) = 40

[elboyran]

Parameters used for MSSRA:

            saliency_types = [1 1 1 1];
            SE_size_factor = 0.02;
            Area_factor = 0.03;
            num_levels = 20;
            thresh = 0.6;
            thresh_type = 's';

Some computed parameters: SE_size = 8 lambda (for DMSR) = 40

[elboyran]

Parameters used for MSSRA:

            saliency_types = [1 1 1 1];
            SE_size_factor = 0.02;
            Area_factor = 0.03;
            num_levels = 20;
            thresh = 0.6;
            thresh_type = 's';

Some computed parameters: SE_size = 8 lambda (for DMSR) = 40

[elboyran]

Committed binary masks for DMSR and MSSRA for the gray_scale image.

[elboyran]

From @dafnevk on May 19, 2016 12:34

How comes that lambda is larger for the DMSR than for binarization?

[elboyran]

From @dafnevk on May 19, 2016 12:55

For MSSRA, I'm missing the parameters for lam_factor and connectivity.

[elboyran]

For the DMSR (and in fact all other detectors, incl, MSSRA) I have looked at the MATLAB code and found out: the binary_detector (the base for all other) have hard codded lam_factor (5) and connectivity (4)! It explains the different lamdba I gave you above. It's not very nice that the code is not parametric, I'll change that (in a new issue),but maybe you can meanwhile test if you get the same results with these values.

[elboyran]

The MATLAB results for MSSRA for the color test image are committed in the Python code repo (in /tests), https://github.com/NLeSC/SalientDetector-python/commit/b8acf2ade7f926f0554dd469847b167049eb0326. For the corresponding parameter and intermediate variable values see the Excel sheet.

[elboyran]

The saleincy masks for the binary detector for the 2 binary test images are added. Now some Python tests fail!!! See also issue #6.

[elboyran]

The MATLAB results for DMSRA for the color test image are committed in the Python code repo (in /tests), https://github.com/NLeSC/SalientDetector-python/commit/6ec60027eb882e1cbc6ff387086df23ea4c9f320 For the corresponding parameter and intermediate variable values see the Excel sheet.

[elboyran]

TO DO: regenerate the DMSRA results for the Gray_scale and color images.

[elboyran]

The result for the color images seems the same: 364 regions. The intermediate variables are SE_size 9 lambda 27 data-driven threshold 125

[elboyran]

The result for the gray-scale image is commited, https://github.com/NLeSC/SalientDetector-python/commit/4a63b335578361adc8c45319059a8eedc27e572b. Now, we obtain 645 regions with SE_size 8 lambda 24 data-driven threshold 142

Excel sheet updated.

[elboyran]

The DMSRA thresholding of the accumulated masks is fixed. New images are generated and Excel sheet updated.