Human and machine similarity judgments in forensic firearm comparisons

Area of Science:

• Forensic Science
• Computer Vision
• Human-Computer Interaction

Background:

• Assessing similarity in firearms comparisons is crucial for forensic identification.
• The performance of human examiners versus automated algorithms in these assessments is not well understood.

Purpose of the Study:

• To compare human similarity judgments with algorithmic assessments in firearms comparisons.
• To determine if humans can accurately distinguish between matches and non-matches in cartridge casing images.

Main Methods:

• Untrained human participants rated the similarity of 40 pairs of cartridge casing 2D-images on a scale of 0 to 100.
• Images were pre-categorized into four similarity groups by an algorithm.

Main Results:

• Humans successfully distinguished between matches and non-matches across different algorithmic similarity groups.
• Human similarity scores outperformed algorithm scores in differentiating matches from non-matches.
• Algorithms demonstrated superior performance in assessing similarity within defined groups compared to humans.

Conclusions:

• Humans are capable of making high-quality similarity judgments in firearm comparisons using two images.
• Human and algorithmic approaches to similarity assessment differ, particularly in group threshold perception.
• A hybrid human-machine approach may enhance the accuracy of firearm identification compared to either method alone.