Naïve Image Segmentation using FloodFill

Posted: February 27, 2012 at 5:30 pm

I’ve been experimenting with using floodFill directly on the morphology output, effectively doing the same thing as my early segmentation approach using mean-shift, but bypassing the mean-shift stage (which was highly computationally intensive). The results are promising:

In an early test of temporal stability these patches are as stable over time as the edge detection, which is far better than the mean-shift stability. There are quite a few more patches per image though, the above image is made of 141 patches, and does not quite reproduce the whole scene, but is still a huge improvement since there are no huge patches that cover the whole image. The current procedure (influenced by the canny segmentation method) is as follows:

  1. Covert the image to greyscale
  2. Equalize (normalize) histogram
  3. Close / open morphology
  4. 3×3 Gaussian blur
  5. FloodFill

I also thought I could use the canny output as a mask to block over flooding, but thus far this is not a problem. How these patches deal with merging and decreasing the amount of data is the next thing to test.