Weekly update for May 29

Bullet matching (Heike)

• scans from Sensofar (Bill Henderson) of bullets by Phoenix PD: 8 barrels (2 shots), 2 unknown bullets
• total of 108 lands (each file ~133 MB)
• available from /myfiles/las/research/csafe/bullet-scans/Phoenix Test/

Changes to bulletr

• changes to parameter settings in bulletr (groove_cutoff, smooth_factor in get_groove both depend on resolution)
• get_groove has now parameter method - just hook any alternative method in here.
• currently available methods are rollapply (Eric’s) and middle (takes middle 75 percent of a land)

Results

• feature set for Phoenix PD bullet lands are a bit strange
• next week: investigate feature set

Chumbley Score

• Chumbley score is non-parametric test based on multiple correlation values extracted from striation marks of two lands (Chumbley et al 2010)

## Warning: 'chumbley' is deprecated.
## Use 'get_chumbley' instead.
## See help("Deprecated")

Chumbley score

# aligned correlation scores
ctest[[3]]

## [1] 0.9188150 0.9362759 0.8998899 0.9380428 0.9556710

# random correlation scores
ctest[[4]]

## [1]  0.20041773  0.82400654 -0.13149010  0.08690511  0.60409433

# results from Mann-Whitney U test
ctest$test

## 
##  Wilcoxon rank sum test
## 
## data:  cor_matched and cor_random
## W = 25, p-value = 0.003968
## alternative hypothesis: true location shift is greater than 0

Chumbley on Phoenix PD (orange is known match)

Chumbley on Phoenix PD, again

Chumbley score

• impact of parameters (window width, number of replicates) on test result?
• variability of results?

More on Grooves (Kiegan and Ulrike)

• working with development version of bulletr, looking at get_grooves
• making sure we understand what the current code is doing
• now with updates for parameter method, we will start looking at new methods to add
• also adding a get_grooves_test that includes plot of smoothed crosscut
• alternatively, an option in the function to output the smoothed plot

Optional smoothed grooves output

• thinking this could be added as an option depending on the method
• allows us to check to make sure the image is “smooth enough”
• if it is misidentifying, this gives a check for whether its because there is a bump before the true peak
• works in my clone of the dev version of bulletr

Original plot from get_grooves

New smoothed option for get_grooves

Handwriting Analysis (Amy)

Two objectives - Begin with human QD examiners specifying characteristics - Automated determination of characteristics

Begin with human QD examiners specifying characteristics

• Sargur N. Srihari, University of Buffalo
• CEDAR-FOX software
• CEDAR dataset
• CEDAR letter
• over 1,500 individuals wrote the letter three times each
• representative of the United States population

CEDAR letter

CEDAR-FOX

CEDAR-FOX

Statistical Models and Inference

• Problem complexity
• Bayesian Networks
• Markov Networks
• Probability of Evidence

Automated determination of characteristics

• Christopher Saunders, PhD
• FLASH ID software (proprietary)
• “… the careful reader should be able to implement a similar system based on the information in this article.”
• Data: modified London Letter (FBI)
• Skeleton of the writing, 1 pixel in width
• Identify nodes and edges
• Topology and Geometry

Graphemes

Shoeprints

• Guillermo and Martin have scans of shoesoles
• Go here to access them
• There are (5 types of shoe)x(2 models with the same outsole pattern)x(left and right feet)x(5 replicates per shoe) = 100 scans.
• We want to ultimately determine whether it is possible to determine whether a shoeprint was generated with a given shoe.
• The parameters of an affine transformation will be find by manually selecting two points on each scan. Code will be available for you in the future (as an R package).