Weekly Update for March 26

Automatically cleaning up shoeprint images

Original	Cleaned Up

Steps: Flatten to 8 shades, remove label, remove background, remove borders, remove straight lines, rotate, crop to lightest columns/rows

REU planning

• 6 offers out. __ accepted
• Working on setting up a site with blogdown
• Lecture & Lunches on Fridays
• “Working Wednesdays” - spend some time with the students improving their professional skills (e.g. CV/resume, cover letter, workplace communication)
• Writing skills - blogposts, Sarah will advise
• Coding skills - weekly R lectures and assignments
• Statistics skills - weekly lectures and assigments corresponding to their projects
• Deliverables: poster session at the end of summer, website with their writing
• Ideas for fun events welcome!

Glass data

• Still just playing around. Get into more this week.

Updates

• Slow Week − Take home exam in progress!
  
• Handwriting Data Collection
  • Target groups of participants ~1500
  • Low cost - send packets, primary communication with a group “leader’’
  • Ask local participants to provide signatures
  • Process:
    • Write prompt in sequence 3 times and simulate signatures
    • Tuesday/Tuesday one week apart (for example)
    • Repeat procedure 3 months later
  • Can you think of any interested groups?
• Project overview for Secret Service

Updates

• NIST scans: Comparing Profiles across all w_o and w_v sizes
• Poster I dont know if this is going to open

SLR’s

• Trying to understand what goes into the construction of an SLR
  • Choosing relevant populations under the prosecution and defense hypotheses
    • Different “anchorings” can result in very different SLR’s (Hepler et al. 2012)
  • Choosing score function
    • Including measure of typicality of the unknown source in the score function itself may be necessary (Morrison and Enzinger 2018)
  • Choosing how to model score distributions
    • Well-known tail problems

SLR’s (cont.)

• Questions
  • What choices are amenable to general guidelines/rules?
  • How can we make SLR’s behave more similarly to LR’s or to each other?
  • Do the above choices depend on each other? For example, could differences in SLR’s be ameliorated by incorporating measures of typicality?

Tests of Matlab’s “imregister” for large images

Tests of Matlab’s “imregister” for large images

Updates

• Temporarily ditching BART
• Adding extremely randomized trees and random rotation forest algorithms
• Should be able to combine extensions to random forests as well
• Image renamer program in CSAFE GitHub organization

Updates

• Working on papers for shoe matching method
• Also working on the efficient number of corners for matching
• Will find other frequently used features for outsole impressions in the imaging analysis
• (Sam and me) Collaboration work with Lund and Iyer on R-package for uncertainty pyramid