Next steps (Skype meeting)
- soft constraints – understand/repair the bug which gives 0s for auxiliary integer variables
- adjacency constraints – implement as soft constraints
- frame fields
- compare with Gurobi
- transition functions as soft constraints
- fit Bezier curves to the sketch via optimal transport: generalization of [DCAD2011], see also video of Pierre’s talk
- Moreover: think about the modification of the greedy solver, specifically how to process adjacency constraints. The constrained points (black pixels with high confidence) could be divided into clusters separated by non-constrained points (white pixels or black pixels with low confidence weights). The clusters could then be processed independently at hte same time.
Optimization terminology
Vertices
constrained
: black pixels, high ST confidenceunconstrained
: black pixels, low ST confidencefree
: white pixels
Constraints
isoline
: snap constrained vertices to uv grid isolinesadjacency
: force adjacent constrained vertices to be close in the uv space (prevent “jumping” to other isolines)
Update
- found the bug which gave all 0s using auxiliary integer variables (
isoline
term) adjacency
term implemented as a soft constraint- the solver now takes a long time to converge (possibly due to a high number of integer variables?)
- even with high weights for the
adjacency
term, the lines are broken: think about how to repair this