It is indeed very interesting!
I surmise there are three contributing factors.
- Your preprocessing steps give rise to large solid white masses.
- There is a follow-on multiplication of the point set scaling all the coordinates up by a factor of 4. In effect, this expands the solid white masses into grids of cities, which you can directly see if you invoke
pointcloud.
again after the multiplication `*. 4’ on the point set image. It would look like this:
- ever-so-slight rounding errors could shift North-South cities closer, perhaps, or vice-versa, and it is likely that such behavior would be the same for large regions. The traveling salesperson would tend in the same direction for long stretches of the path until local errors in plotting shift to favor other directions, such as the transition diagonals in the upper left hand patch of the tiger’s head.
You might — for its entertainment value, substitute for *. 4
a fill command f. c==0?i*4.1:i*3.9
to bring cities closer together slightly along the North-South axis; this would lead to a preponderance of vertical lines, like so:
as the traveling salesperson finds North-South cities slightly closer in the grid than East-West cities. Overall, there are a number of interesting tricks to be pulled here.
@lylejk
Thank you for the pointers! However, these also re-enforce my thinking (perhaps not clear) regarding what I think are good gmic-qt
filters: few controls and large “sweet spots” that accommodate many kinds of images, and such that one does not have to do much preprocessing either in preparation or internally, through a number of pre-tsp
image conditioners, in order to get splendid results. If a lot of preparation needs to be done in order for an image to be in the filter’s ‘sweet spot’, then the likelihood of the filter user being disappointed is that much greater, and that is especially an important matter of concern for compute-intensive, long-running filters. After prep and long minutes of waiting, the results really need to be good or people will be broken-hearted.
I really like the TSP tool you are using; it responds to tonal changes quite well. It would be interesting to take it apart and see how it ticks. Alas! I am over-booked and see no time for such an exercise on my watch. Perhaps, in a week or two, if some interesting mods show up here, I could wrap it up as a gmic-qt
filter, but I make no promises. I also have an inventory of tutorial material nearing release (there has been none from me since last September), and that sits higher on the TODO stack.
And — there’s the rest of life, too…