I wondered, because corresponding to your original image (the 2px particles), there is no blurry mish-mash like with regular gauss- or box-blurs. I'm just asking, because in the long run I'm looking for better and faster alternative to blur my SSAO images, the better if the kernel is dynamically sizeable
I think I have a run at the moment; instead of playing Skyrim I coded a contour tracing and region segmentation algorithm in Lite-C. Even though I didn't fought against dragons and took some Skoma, it was quiet fun, too:
The changing color in the contour image indicates the order of contour points, stored in a linked list. It can handle degenerated cases, too (1px lines).