[Work Log] Work Log
January 22, 2015
Next up: finding trees in single image.
- User specifies number of branches
- Prior over branch width in pixels
- Weak prior over length (with hard bounds)
- Don't do anything clever
Tasks
- Train new likelihood forcing curves to be constant-width.
- Generate proposals from skeletonized foreground.
- Optimize likelihood of proposal by extending/shortening.
- Extend, find MAP, estimate ML Assume isotropic likelihood hessian Known prior hessian ** Laplace approximation
Skeleton to proposal
- pick long chains (weight by length). Extend by connecting chains if below angle threshold
- convert to medial axis form, place in image
Evaluate proposal
- optimize posterior
- compute ml
- maximize ml by repeatedly extending, optimizing, computing ml
- maximize ml by repeatedly shortening, optimizing, computing ml
propose branching
- optimize Laplace-approximated ml over (a) branch distance; (b) branch position