Triangulation

Triangulation based on those found tracing lines.

My aim is to find a segmentation based on those tracing lines.Also I am trying to find critical points based on Greene's bisection method. But right now it seems not working.

A promising vector field

Modified fields

Problems for which directions should be picked?

Questions?

1. Still the smoothed corners(Should do something in tracing)
2. Swirling(Use salient directions or explicitly point out the situation?)
3. The choice of stopping points for a line; The long strokes are not always the best choice(see the Lena result). Small stokes prefers in some detailed regions or textures. Long strokes are good for the silhouettes of an object.
Two different hysteresis for threshold did control the length. Are there any easy or automatic way to choose the best parameters for different cases. By the way, this idea is used by Kang's paper.
(Should summary some rules based on perceptions or drawing techniques.)
4. How to get crossed lines?(Should have a better representation for vector field? )
5. Not good for keeping small details.Cannot show gentle structures.

Images of ETF after fDOG

Some testing

Maximum length = 20, Density control after finding a line by refusing putting a seed in the neighbourhood

Bigger images

Without length control

Bilateral smoothing by the directional differences

Drawing only pixels at ETF edges

As for the testing for bilateral smoothing, the angle similarity between pixels did affect the results a lot. Otherwise, the intensity, the distance only has a little influence.

 

Not much difference between previous smoothing and bilateral smoothing

There is very little change in very small regions. As for the reason, I guess it is because the LIC integrates the difference.

Previous smoothing considering only gradient magnitude difference

Bilateral smoothing considering both position and gradient magnitude difference

(sigmad=12, sigmar=6)

Visualize vector field by LIC