November 05, 2014

Met with Kobus and Andrew regarding the seed counting task. Andrew summarized the current state of the project. We have foreground-background segmentation in octave, and it seems to work well in the easy cases. If I understand correctly he has some code to match possible exemplars with an example exemplar, using an affine transform to try to align them. Given this set of aligned exemplars, a shape model could be learned for more sophisticated analysis.

Kobus outline a strategy for generating a "heatmap" for data-driven proposals:

- find edgelets
- for each edgelet, try matching to the contour of an exemplar. If it matches well, add a "vote" to the pixel where that exemplar wouid appear (or rather it's centroid). Could give partial votes for partial matches. Could give weaker votes to shorter edges.

Possible things to show for 10m talk:

- pipeline: thresholding, interactive edge extraction tool.
- a sketch of the algorithm
- preliminary results

Talk outline:

- outlining the problem, show example image
- its difficulties, show example image,
- what has been done, show thresholding,
- motivate planned work: exemplar model implies a large search space
- what is planned to be done, and motivation for that plan.
- current progress

Summarizing the "motivation" step: searching the entire image over the entire space shape (including the number of shapes) is an exponentially growing problem. Need a set of good candidates. Our "exemplar map" could be a good way to reduce the search space by only considering locations where the seed is likely to appear.

Andrew provided some resources that could help

- slides from earlier presentation are in: /space/predoehl/iplant-demo-2012oct17.tar.bz2
- octave code for thresholding is in /data/predoehl/seed-counting/.../...

To do before CVPR is over:

- run edgelet extraction on all images

After CVPR:

- implement voting scheme, run on example image`

Other thoughts

- Could my edge-based likelihood be useful here? :w