We tested both the snake and the region growing algorithm on a synthetic image sequence of a circle changing shape. We also manually drew a boundary as we would have done for images without a specified ground truth to check the accuracy of the eye. For all results, red pixels are correctly segmented, green pixels are over-segmented, and blue is undersegmented. Note that vbndcmp sometimes colours the correctly segmented portion of the first image in the sequence white instead of red in some cases.
Region Growing
Active Contour Model
Accuracy (by Fraction Correct Pixel)
Hand: 0.951
Region Growing: 1.000
Snake: 0.928
--------------------------------------------------------------------------
Circle with Noise
We added noise to the previous image and tested region growing and snakes on resulting image. For the ground truth, we used the same ground truth as the image without noise.
Region Growing
Active Contour Model
Accuracy (by Fraction Correct Pixel)
Region Growing: 0.699
Snake: 0.819