Europhlukes To Do List

18 November 2003

1. To get test data for humpback matching. (contact Michael Kingsley)
2. To improve the threshold- globally and locally. Try fitting some sort of average gray-value surface and compute deviation wrt this surface (data-dependent local threshold).
3. To develop appropriate matching algorithm. We should at least be able to automatically generate the "standard codes":
4. To submit to the ECS conference (ext. abstract till 1 Dec. 03) if progress. (done)
5. To investigate how to make it MATLAB independent (compilation, Java, etc.).
6. To adapt the program for other species like killer whales and bottlenose dolphins.
7. To investigate image quality measures. Have a look at Gemma's thesis, in particular, paper by Wang and Bovik.

13 Jan 2004

• Other shape measures/descriptors
• Icip paper: end of Jan
• Medial axis is not sufficient, we need to estimate a model whenever the leading edge is submerged.
• Shape descriptors:
• angular radial transform (region-based)
• turning angle
• elastic deformations
• Fourier descriptors,
• curvature scale space
• A completely different approach could be based on "points of interest": (eg. via SIFT, Scale-invariant Feature Transform, D. Lowe) automatically compute a large amount of interest points in the image, restrict attention to tail area, compute their spatial distribution in an affine invariant way and match.
• Advantages:
• No real segmentation necessary
• relatively robust wrt global contrast changes, etc, really looks at patterns
• Robust wrt occlusion

2 February 2004

• Is this construction really affine invariant?? In a really affine invariant approach one would preferentially start drawing the vertical lines, by connecting points with the same relative length coordinates along the trailing and leading edge of the fluke: ie. midpoints to midpoint, quarter point to quarter point etc. The horizontal line would then be obtained by slicing through the vertical lines.
• Saliency and key-point detectors: doesn't look very promising at the moment.
• Larger testbase (contact Americans)
• What will we do about the submerged regions.
• Use features that are based not just on area but on more sophisticated measures.

9 February 2004

• Contact Americans!!! Done!
• Affine invariant construction of midline.

23 February 2004

• New grid, any results;
• scratch detector based on morphology
• slide for ECS conference

4 March 2004

• Edit tech reports for publication and dissemination at ECS

22 April 2004

• Irish Vision Conference, submission? Probably not possible!
• Trainee? Using Matlab compiler: Elena will write short "stageplan";
• Student input from Judy Allen side: preferably after visit by Elena? When would be a good time?
• Documentation for the different procedures/ if possible on the web. XML-based.
• Plans for further development?
• Manual adjustment of tail
• Detecting and characterizing salient bits?
• Matlab library effort: where and how do we start?