class: center, middle, inverse, title-slide # Update Jul 8 2019 ### 2019-07-08 --- class: inverse # Ballistics --- class: primary # Overview Faculty - Heike Hofmann - Susan VanderPlas Graduate Students - Ganesh Krishnan - Kiegan Rice (& Nate Garton) - Charlotte Roigers - (Joe Zemmels) Undergraduates - Talen Fisher (fix3p) - Mya Fisher, Allison Mark, Connor Hergenreter, Carley McConnell, Anyesha Ray (scanner) - Molly McDermott, Andrew Maloney (REU-CSAFE) - Syema, Tiger, Emmanuelle (REU-SPRITE) --- class: primary # Bullet projects - Big picture - **data collection** - **computational tools** - matching lands: 1. crosscut identification 2. **groove location** 3. curvature removal 4. alignment of signatures 5. feature extraction 6. matching with trained Random Forest - **analysis of results** - **communication of results** --- class: primary # Update from the data collection - **scans from bullet lands (about 20,000 total)** - LAPD: 4 bullets per barrel for **all** of 626 firearms - Hamby Sets 10, 36, 44, 224, and a clone (35 bullets each) - Houston test sets (6 kits with 25 bullets each) - Houston persistence: 8 barrels with 40 fired bullets each - St Louis persistence: 2 barrels with 192 fired bullets each - most of the CSAFE persistence study - **and cartridge cases** - DFSC (about 2000) - getting ready to scan cartridges for CSAFE persistence - **shooting range** - we are planning to go out to the range *one* more time to finish up the persistence study --- class: primary # NEWs - visit to Vic Murillo from DCI Ankeny  --- class: primary # Charlotte: groove detection - **Goal:** To automate the process of identifying groove engraved areas of bullet scans using computer vision algorithms. - **Process:** - Prep images using gradient thesholding - Perform a Hough Transform on edges to identify groove edges - Find Hough lines closest to the middle 2/3rds of the bullet scan at the top and bottom of the bullet scan .center[ <img src = "charlotte/07_08_2019_Update/process_descrip.png" width = "49%" /> ] --- class: primary # Charlotte: groove detection .center[ <img src = "charlotte/07_08_2019_Update/Houston_BarrelF_Bullet2_Old_Process.png" width = "55%" align = "top"/> <img src = "charlotte/07_08_2019_Update/Houston_BarrelF_Bullet2_New_Process.png" width = "55%" align = "bottom"/> ] --- class: primary # Charlotte: groove detection .center[ <img src = "charlotte/07_08_2019_Update/sad.compare.png" /> ] --- class: primary # Ganesh: Diagnostics in the bullet matching pipeline ### *Shiny Applications* - **Interactive user interface for performing transformations, preliminary evaluations, extraction and scoring, and batch operations** - **Diagnostics using Interactive visualizations in the bullet matching pipeline** .center[ ### Demo ] --- class: primary # Kiegan: variability study <img src = "kiegan/scanning-stage3.png" width="45%"/> --- class: primary # Kiegan: variability study Collected ~2000 scans: - 9 bullets, 3 each from 3 barrels - 6 LEAs per bullet - 5 operators - 2 machines - 3+ repetitions of each environmental condition --- class: primary # Kiegan: variability study Example of all 90 repetitions for a single signature. <img src = "kiegan/bo_land3_all.png" width="100%"/> --- class: primary # Kiegan: variability study Initial results show: - Differences between barrel types - Deeper striations = larger machine effect - More tank rash/breakoff = larger bullet effect - Lots of structure due to bullet - Each bullet marks slightly differently - Measured "peaks and valleys" differ --- class: primary # Kiegan: variability study <img src = "kiegan/bo_land3_bullet.png" width="100%"/> --- class: primary # Kiegan: variability study Currently working on modeling: - At the signature level - Using all X locations, removing overall structure - Subsampling to each 100th X location - In 5 different phases to check for consistency - model 1: (1, 101, 201,...) - model 2: (21, 121, 221,...), etc. - Pairwise scores - Same-source pairwise scores - Different-source pairwise scores - For completeness - Balanced with 3 repetitions - Hierarchically with 3-5 repetitions - With/without one operator --- class: primary # Update on REU projects REU-Sprite: assessment of scan quality <img src = "heike/variability.png" width="100%"/> --- class: primary # REU-CSAFE **Goal:** Validate `bulletxtrctr` features in comparison to `bulletr` features used to fit the random forest model **Process:** - Get Hamby 173 and 252 scans from NIST - Process scans to get new features using `bulletxtrctr` - Match new features to old features - what changed? Currently, trying to match features from Hare et al. (2016) to new features --- class: inverse # Glass --- class: inverse # Soyoung --- class: primary # Glass project overview - Comparison of claasifiers of standard method (ASTME2330) and the RF method - Data collection: from two companies, collect sample of panes consecutively manufactured <img src = "soyoung/box_plot.PNG" width="80%"> --- class: primary # Glass project overview - Sample of data: 18 chemical compositions were measured <img src = "soyoung/sample.PNG" width="70%"> --- class: primary # Glass project overview - Density estimations <img src = "soyoung/density2-1.PNG" width="70%"> --- class: primary # Glass project overview - Correlation among 18 elements <img src = "soyoung/corr.PNG" width="60%"> --- class: primary # Glass project overview - ASTM-E2330-12, ASTM-E2927-16 - Construct the interval on each element and test on the individual element - Use threshold 4, to declare mates/non-mates <img src = "soyoung/ASTM.PNG" width="70%"> --- class: primary # Glass project overview - The random forest(RF) method - Train the RF to understand features for mates and non-mates <img src = "soyoung/simpletree.png" width="50%"> --- class: primary # Glass project overview - Source prediction - 69120 comparisons <img src = "soyoung/ROC.PNG" width="70%"> --- class: primary # Glass project overview - Source prediction - 69120 comparisons <img src = "soyoung/performance.PNG" width="80%"> --- class: primary # Shoes --- class: inverse # Miranda --- class: primary ## CoNNOR: Convolutional Neural Network for Outsole Recognition - Goal: Recognize bowtie, chevron, circle, lines, polygons, quadrilaterals, stars, text, and triangles in outsole images - Input data: labeled geometric shapes on outsole images - \> 80K shoe outsole images from Zappos.com - ~ 5K have been labeled with one or more polygons - ~28K polygons total.  --- class: primary ## CoNNOR: Convolutional Neural Network for Outsole Recognition - Model is relatively accurate, but fails in some interesting ways...   - Paper describing CoNNOR submitted to FSI last week! --- class: primary ## In other news... -Written prelim starts one week from tomorrow  --- class: primary # Susan - `ShoeScrapeR` package: obtain photos of shoes from Zappos.com with appropriate metadata - Currently updating to obtain metadata and images of side, top, bottom, back, front of the shoe, along with brand logos - mostly done, but still working to make it robust enough to run on it's own - optimizing to reduce repeated downloads of the same data --- class: inverse # Eryn --- class: primary # Project Updates **Goal of Project** To measure the wear pattern of a single shoe over time - taking the longitudinal study's 3D scans of soles of shoes .center[ <img src = "Eryn/Nike_3d_bottom.png" width = "70%" /> ] --- class: primary # Project Updates Continued - Alignment posible solultions - Manual landmark: find specific areas of the shoe to mark as a landmark - length detection: finding the longest and shortest length of the outline of the shoe as initial alignment .center[ <img src = "Eryn/heirNike.png" width = "70%" /> ] --- class: primary # Statistical Foundations --- class: inverse # Danica --- class: primary # Statistical Foundations Overview ISU Faculty - Danica Ommen - Yumou Qiu ISU Students - Nate Garton --- class: primary # Project Updates Statistical Foundations - PI on NIJ Grant: Error Rates, ROC Curves, and the Two-Stage Approach - Initial results will be presented at SimStat in Sept. - Difference between Bayes Factors and Likelihood Ratios - 1 publication submitted to Statistical Science (waiting on reviews) - 1 draft publication with Soyoung to be submitted to Law, Probability, and Risk - NSF Proposal: Likelihood framework for Forensic Evidence - Nate Garton: Validity of SLRs for Forensic Evidence - 1 draft publication in the works --- class: primary # Handwriting --- class: primary # Handwriting Overview ISU Faculty - Danica Ommen - Alicia Carriquiry ISU Students - Nick Berry (former student, currently at Berry Consultants) - Amy Crawford - James Kruse --- class: primary # Project Updates Handwriting - NIJ Grant: Validity of FDE Determinations using Kinematics - Collected writing from 33 different writers (static and dynamic) - Sent a subset of static writing pairs to 40 examiners - Examiners asked for strength of support (1-7) for same writer (Q1) or different writer (Q2) - Is there a relationship between dyanamic features and examiner support? - Initial results will be presented at ASQDE in Aug. - Amy Crawford: Bayesian Hierarchical Models for Writership of Handwriting - Inital results to be submitted soon (with Nick) - 1 draft publication coming soon --- class: inverse # James --- class: primary # Data collection Purpose: - Data is to be used in an open access database and to test materials developed by CSAFE researchers Our Goal: - To enroll 100 participants - Have all sessions completed by summer's end - Another batch of participants will be run in the fall Currently: - 125 Participants --- class: primary # Data Collection: Continued Collection: - Three Collections <SUP>&nbsp;</SUP> - First in person/video/phone - Need a group of at least three to participate at a distance. <br /> <br /> - Survey - Three paragraphs - three reps of each - Sign a fake name <br /> <br /> - Samples are scanned once completed --- class: inverse # LateBreak --- class: primary # Late Break News --- class: inverse # Issues --- class: secondary - [Issues!!](https://github.com/CSAFE-ISU/slides/issues) - One issue down, three to go.