Image Acquisition and Comparison.

The TopMatch matching algorithms are completing beta testing. They will continue to be improved through 2016 and all TopMatch systems will receive free software updates. Parts of the system described below are patent pending.

  • 3D Scan Acquisition is based on GelSight imaging technology that uses an elastomeric sensor and enhanced photometric stereo.
  • Scan acquisition is fast, requiring approximately 2 minutes per scan.
  • Our novel casing comparison algorithm is a feature based approach, which automatically locates and matches the same types of marks that a human examiner would identify.
  • The true breech-face impression and the aperture (primer) shear are compared separately. A unified match score quantifies the degree of similarity between these two toolmarks for a pair of casings.
  • A newly developed statistical scoring function complements the underlying match score to provide a statistical basis for each hit.


3D Scan Acquisition

The TopMatch system utilizes advanced three-dimensional imaging algorithms (e.g., shape from shading and photometric stereo) and the retrographic sensor of Johnson and Adelson [Johnson, Adelson, CVPR 2009; Johnson, Cole, Raj, Adelson, SIGGRAPH 2011] (GelSight) to measure an object's three dimensional surface topography. The retrographic sensor is a block of clear elastomer with a thin layer of elastic paint on one side. When an object is pressed into the elastomer, the layer of paint conforms to the shape of the surface. The layer of paint removes the influence of the optical properties of the surface on shape estimation. Every material, such as glass, metal, plastic, or human skin, appears the same when pressed into the retrographic sensor. In contrast to confocal microscopy and focus-variation microscopy, this important feature of our system removes the influence of surface reflectivity on the measured topography. Although firearms examiners typically consider toolmarks ranging from tens to hundreds of microns in diameter, they do occasionally consider features as small as 10-15 microns (see figure below). The scanning resolution of the TopMatch-GS 3D system is capable of capturing and representing these small marks.


A 3D scan of the NIST Standard Casing (SRM 2461) illustrating several distinct features with their approximate sizes. Scan at 1.4 micons/pixel. Firearms examiners often rely on features 10 microns and larger.


The elastomeric sensor is mounted on a sheet of glass and a camera views the reflective skin through the clear gel (see figure below). A set of lights sequentially illuminates the sensor to reveal an initially shaded image. Calibration images of a grid of spheres are used to build a photometric model relating image intensity to surface orientation. Images of the object pressed into the sensor are collected and converted to surface normals using nonlinear least-squares optimization. A surface normal is the vector perpendicular to the object surface at the specified (x, y) location. The surface normals are then integrated to obtain a three-dimensional surface. The scalar value recorded at each pixel is the surface height of the object at the corresponding location. A particular strength of the GelSight imaging technology is its ability to capture surfaces with significant slope. This provides TopMatch an advantage over confocal microscopy whose signal can become unreliable for slopes greater than 15-degrees. In contrast to other 3D imaging technologies which suffer from pixel drop-outs (i.e., pixels with missing depth measurements), our GelSight based scanner measures a complete surface without missing measurements.


(Left) Illuminated lightplate with a circular piece of gel located under the glass. The small impression of a cartridge casing is just visible in the center of the gel. (Right) Closeup of casing under the gel.


Although the TopMatch-GS 3D system does not support fingerprint analysis; we created a custom finger holder and collected this demonstration scan. The scan was acquired at a resolution of 4.2 microns/pixel. Significant detail, including sweat pores, can be seen. Note that the 3D scan was acquired off living tissue.


Matching Algorithm

The TopMatch matching algorithm is based on ideas from the Computer Science subfield of computer vision. Automatically identified distinctive features are used to match and align two casings. Our algorithm identifies a maximal set of self-consistent matched features which correspond to geometric parts of traditional toolmarks. A set of matches is considered self-consistent if the matched features of two casings can be spatially aligned after a single rotation and translation of one image (modulo the scale). The score of the match is a function of the number and quality of matched features.

By requiring spatial coherence of matched features, the methodology is able to strongly indicate when two casings were fired through the same firearm. In contrast to cross correlation based methods, feature-based techniques compute the match score using only the portions of the image identified as informative (i.e., the features). Features correspond to regions of the image with nonzero gradients in both the x and y dimensions. In other words, the method looks for corner-like textured regions of the surface, of any size, which correspond to the same types of ridges, peaks, gouges, and concavities that a trained firearms examiner would identify. Because the scanned surface is the true 3D shape of the surface, it is not necessary to cope with appearance variations (e.g., shadows) that arise due to variations in lighting conditions.

A statistical scoring function complements the raw match score. This statistical scoring function can assign a more interpretable probability of match to each candidate comparison. There are two main motivations to developing the statistical scoring function. First, we wish to begin putting breach-face impression matching on a firmer scientific footing by associating a statistically meaningful probability of match. Second, we want to improve the quality of the match ranking provided with this function, as it is able to account for not only the number of matched features but other quantities that measure the quality of those matches. For example, the number of matched features, the average difference in feature appearance, differences in feature scale, the size of the masked region, and the overall fraction of the masked region covered by matched features. We believe that the current model is a bit conservative. That is, at the 1 in 10,000 (99.99%) confidence level we might expect 20 of 200,000 known non-matches to appear significant (i.e., a false positive rate of 1 in 10,000); however, we see no false positives in 200,000 comparisons.


The masked region of the breech-face impression appears in color, as a painted surface, in the 3D viewer. Because the masking takes place in the 3D viewer, the user can zoom and rotate the surface in three-dimensions.


The current matching algorithm compares the true breech-face impression and aperture shear (see figure below). It does not consider the firing pin impression, the ejector mark, or the extractor mark. During scan acquisition, an auto-masking algorithm generates an initial estimate of the location of the breech-face impression which can be refined by the examiner.


The linear profile of two extracted aperture shears are shown for a pair of casings fired through a Norinco firearm.


The ability to compare aperture shear linear profiles enhances the system's ability to match casings from firearms that produce strong shears (e.g. Glocks). Throughout 2016 we will be improving both our breech-face impression and aperture shear matching algorithms. All algorithmic improvements will be provided to all actively deployed TopMatch systems.


Heatmap Visualization (patent pending)

Unlike other system which provide little to no explanation on the detail of a match, the TopMatch system can display a heatmap to explain the computed match score. With a heatmap, the software color codes the surface of the casings to indicate the areas of geometric similarity (patent pending). This visualization provides interpretability and facilitates communication of your findings to those who are not experts in firearm forensics.

A known-match pair of 9mm Luger A known-match pair of 9mm Luger casings. Color intensity corresponds to the number (and density) of matched features. Darkly shaded regions have a higher degree of similar surface geometry than lightly shaded (or unshaded) regions.

Next: Hardware and Software Demo Video

Cadre Forensics is the forensics division of Cadre Research Labs, a Scientific Computing Contract Research Group specializing in Algorithm Development and Technology Transfer for interdisciplinary applications. Leveraging both internal and university sourced research, Cadre goes beyond off-the-shelf software to develop the next generation of research and discovery tools. We specialize in utilizing techniques from Machine Learning, Algorithm Development, Data Mining, and Computational Biology to tackle both custom computational challenges and the analysis of 'Big Data'.

Cadre Research

Location:Chicago, IL USA Phone:(312) 620 9958 Web: