Presentation part of the course! This starts in 3 weeks Pick a research paper related to image analysis and present it It's ok if it's not on biological images! Pick a paper at least a week before presenting Google Scholar can help find one We've got a lot of people, so you'll only have about 18 minutes Fixes for last time's example, demo with 2D working Cheat information: Threshold of 160 mostly works To figure this out: Print out pixel under event Wrapping around screen? Leaks and incomplete finding: Both happen! It would be nice to see the leaks without waiting 3 hours! Doing this outside the UI thread 3D use: This takes a while When I tried it, there was a leak, which made it take forever Maybe 3 hours or so! Slowness in 3D: We've got two major problems: Not multithreading Python And one opportunity: The GPU This problem is not "embarrassingly parallel" We're using a shared data structure It *might* be thread-safe, but it also might not! A really simple idea: Each thread maintains it's own queue Initially, we start the 6 or 26 threads with separate queues Queues may have duplication, so we check color on entry AND exit CPU time loss from duplication will be less than for mutex, etc Avoids memory hotspots Leak problem 1: Boundary detection Just picking a threshold isn't really good enough Ideas for boundary detection: 1. Set up an elaborate AI boundary detector This has worked in the past, with neural nets and such 2. Highlight the boundaries with a simple technique For example, subtract change in intensity to darken them 3. Tolerate the leaks Bicycles and slime We've tried a lot of other ideas here For example, using linear regression to follow boundaries This worked pretty good Leak problem 2: Stack alignment Demo with up/down on misaligned slice Ideas for alignment problem: 1. Manually move or remove bad sections But lots of sections are a little bit off! 2. Run an alignment algorithm That can work 3. Remove/edit really bad slices, align others For publication, must be very clear about this process!