Scalable spatial point process models for forensic footwear analysis

1. Problem Statement

Forensic footwear analysis relies heavily on "accidentals" (Randomly Acquired Characteristics or RACs)—unique cuts, scrapes, and wear patterns on shoe soles—to distinguish between shoes of the same make, model, and size. While thousands of identical shoes are manufactured, accidentals can theoretically link a specific suspect's shoe to a crime scene print.

The core statistical challenge is quantifying the Random Match Probability (RMP): the probability that a randomly selected shoe from a relevant population would produce a print with accidentals consistent with a crime scene print, given that the class characteristics (size, make, model) and general wear patterns already match.

Key Limitations of Existing Methods:

• Data Scarcity: Datasets like JESA (386 shoes) and WVU (1,300 shoes) are convenience samples with diverse designs but few repeats of the exact same shoe model.
• Model Scalability: Previous Bayesian hierarchical models (e.g., Spencer and Murray, 2020) were computationally expensive, limiting their application to larger datasets like WVU.
• Lack of Specificity: Many existing methods aggregate data into "heat maps" that ignore the specific contact surface (tread pattern) of an individual shoe, failing to model the relationship between the shoe's specific geometry and where accidentals are likely to occur.
• Uniformity Assumptions: Early models assumed accidentals were uniformly distributed, which empirical data has proven false and leads to overconfident (conservative) conclusions.

2. Methodology

The authors propose a Hierarchical Bayesian Latent Gaussian Spatial Point Process Model designed to be both expressive and computationally scalable.

A. Data Representation

• Input: The model uses two primary inputs for each shoe $s$:
  1. Contact Surface ($C_s$): A high-resolution grayscale image of the shoe's outsole (tread pattern), processed into a $39 \times 91$ grid. Darker pixels indicate contact areas.
  2. Accidental Locations ($Y_s$): Coordinates of accidentals on the shoe, aggregated into counts per grid cell.
• Preprocessing: Images are reflected to align left/right shoes, cropped to the shoe boundary, and coarsened from $869 \times 869$ to $39 \times 91$ pixels to reduce dimensionality while retaining tread details.
• Covariates: The model incorporates:
  • Pixel intensity (contact surface).
  • Image gradient (Sobel transformation) to capture edges.
  • Neighborhood interactions (local contact patterns).

B. Model Formulation

The accidental counts $Y_{s,a}$ at grid location $a$ for shoe $s$ are modeled as a Poisson process:
$$Y_{s,a} | C_s \sim \text{Poisson}(\lambda_{s,a})$$

The log-intensity $\log(\lambda_{s,a})$ is modeled as a Latent Gaussian Model (LGM):
$$\log(\lambda_{s,a}) = \beta_{\text{smooth}, a} + \beta_{\text{shoe}, s} + \sum_{i \in I} x_{s,a}^i \beta_f^i + \sum_{i \in I^*} x_{s,a}^i \beta_{sv, a}^i$$

Where:

• $\beta_{\text{smooth}, a}$: A global spatial trend common to all shoes.
• $\beta_{\text{shoe}, s}$: A shoe-specific random intercept (nuisance parameter) accounting for total accidental count variability.
• $x_{s,a}^i$: Interaction terms involving contact surface intensity, image gradient, and their neighbors.
• $\beta_f^i$: Fixed effects (global coefficients) for interaction terms.
• $\beta_{sv, a}^i$: Spatially varying coefficients allowing the effect of covariates (e.g., how much a cut is likely to form on an edge) to change depending on the location on the shoe.

C. Inference Strategy

• Challenge: Standard Markov Chain Monte Carlo (MCMC) is computationally infeasible for the WVU dataset ($1300 \text{ shoes} \times 3549 \text{ grid points} \approx 4.6 \times 10^6$ data points) due to the high dimensionality of parameters.
• Solution: The authors utilize Integrated Nested Laplace Approximations (INLA). By framing the problem as a Latent Gaussian Model, they can perform fast, deterministic approximate Bayesian inference. This allows for the estimation of posterior distributions for thousands of parameters in a fraction of the time required by MCMC.

3. Key Contributions

1. Scalability via INLA: The primary contribution is reframing the forensic footwear problem into a Latent Gaussian framework, enabling the use of INLA. This allows the model to scale to the large WVU database (1,300 shoes), which was previously computationally prohibitive for complex hierarchical Bayesian models.
2. Continuous Contact Surface Modeling: Unlike previous work that binarized the tread pattern (contact vs. non-contact), this model uses the continuous grayscale intensity of the contact surface, preserving more information about the shoe's geometry.
3. Spatially Varying Coefficients: The model introduces spatially varying coefficients for covariates (specifically pixel intensity and image gradient). This captures the non-linear, location-dependent relationship between the shoe's tread features and the likelihood of accidental formation (e.g., edges might be prone to cuts in the toe area but not the heel).
4. Image Gradient Integration: The inclusion of the image gradient (Sobel transform) as a covariate allows the model to explicitly account for edge features, which are critical for accidental formation.

4. Results and Evaluation

The model was evaluated using 10-fold cross-validation on the WVU database, comparing the proposed method ("Our Method") against several baselines:

• Uniform Model: Assumes random distribution (Stone, 2006).
• Spatial Smoothing Only: Aggregates data without specific shoe features (Yekutieli et al., 2012).
• Binary Contact Surface: Uses a thresholded contact surface without spatially varying coefficients (Spencer and Murray, 2020).
• Variants: Models with different combinations of fixed vs. spatially varying coefficients.

Performance Metrics:
The evaluation used a log-probability metric based on the multinomial component of the accidental distribution (focusing on spatial configuration rather than total count).

• Superior Predictive Accuracy: The proposed method (M(Final)) achieved the lowest log-loss (best performance) in all 10 folds, outperforming the binary contact surface model and all other baselines.
• Robustness: The proposed model showed superior robustness in the "lower tail" (shoes with poor fit in other models), which is critical for forensic reliability.
• Feature Importance:
  • Models excluding contact surface information performed significantly worse.
  • The inclusion of image gradient terms provided a measurable improvement over models that only used pixel intensity.
  • The "Median Performance Loss Ratio" showed that the proposed method was roughly 100 times more accurate than the binary contact surface model and orders of magnitude better than the uniform model.

Interpretability:

• The model identified that accidentals are less likely in the arch and toe/heel tips, and more likely on the inner heel and sides of the toebox.
• Crucially, the model confirmed that higher contact surface intensity (darker pixels) and stronger image gradients (edges) are positively correlated with higher accidental propensity.

5. Significance and Conclusion

This work represents a significant advancement in forensic statistics by bridging the gap between complex Bayesian modeling and large-scale forensic data.

• Forensic Impact: By providing a more accurate and data-driven estimate of Random Match Probabilities, the model helps forensic examiners avoid over- or under-estimating the strength of evidence. It moves the field away from subjective expert opinion toward empirically validated, probabilistic statements.
• Methodological Impact: The successful application of INLA to spatial point processes with spatially varying coefficients demonstrates a pathway for scaling complex hierarchical models in other forensic and spatial domains.
• Future Directions: The authors note that while the model works well on high-quality lab data, future work must address the translation of these models to partial, noisy crime scene prints. Extensions to non-linear effects (GAMs) and alternative registration schemes are also suggested.

In summary, the paper presents a scalable, interpretable, and highly accurate Bayesian framework that leverages the full geometric detail of shoe prints to quantify the rarity of accidental patterns, thereby strengthening the scientific foundation of forensic footwear analysis.