Following our yesterday’s Skype meeting (Adrien, Misha, Tibor) below is a short summary of the current state of the vectorization project.

We still haven’t abandoned the idea of submitting to Eurographics, although the deadline (Tue Oct 9) looks a bit tight. I’ll do as much as I can in the next two weeks, then we can make a final decision.

To recap, here’s the pipeline:

  1. Estimation of tangent constraints from the sketch
  2. Computation of the frame field
  3. Computation of the parametrization
  4. Extraction of integer isolines and fitting of Bezier curves

Step 4 is the last missing piece. When that’s done, I can focus on improving the individual parts.

Results & comparison

We’re hoping the overall results will be comparable to Misha and Justin’s paper. Additionally, we will have:

  • local feature-size control via a scalar field (computed automatically or specified manually)

screenshots/Screenshot-2018-09-11-input.png screenshots/Screenshot-2018-09-11-gradient.png screenshots/Screenshot-2018-09-11-result.png

Possible applications

  • computing line drawings from bitmap images, similarly to [Kang et al. 2007]: apply edge detector, then our method, use simple images with few curves to demonstrate this

Extraction

screenshots/IMG_20180906_182147.png

UV-labelling

Tangent constraints are now computed using Sobel, and not structure tensor + bilateral filter as before. This means that the constrained frame field direction cannot be determined by finding frame root closest to the tangent direction (too much noise).

For now, the solution is to use Sobel for frame field constraints, and Structure tensor for uv-labelling. If we have some time in future, we could try to solve this using a global approach such as graphcut.