Scalable spatial point process models for forensic footwear analysis

Alokesh Manna; Dipak K. Dey; Neil Spencer

arxiv: 2602.07006 · v2 · submitted 2026-01-30 · 💻 cs.CV · cs.LG· stat.ML

Scalable spatial point process models for forensic footwear analysis

Alokesh Manna , Neil Spencer , Dipak K. Dey This is my paper

Pith reviewed 2026-05-16 10:03 UTC · model grok-4.3

classification 💻 cs.CV cs.LGstat.ML

keywords forensic footwear analysisspatial point processeslatent Gaussian modelsINLAshoe print accidentalsBayesian hierarchical modelsspatially varying coefficientstread pattern modeling

0 comments

The pith

A latent Gaussian spatial point process model with spatially varying coefficients, using INLA, enables scalable analysis of accidental marks on shoe prints.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a hierarchical Bayesian model to quantify how rare particular patterns of cuts and scrapes are on recovered shoe prints. It reframes the locations of these accidental marks as a latent Gaussian spatial point process so that integrated nested Laplace approximations can perform inference on large collections of annotated prints. Spatially varying coefficients are introduced to let the underlying tread pattern influence where accidentals are more likely to appear. The resulting model shows better predictive performance on held-out shoe prints than earlier approaches.

Core claim

The central claim is that accidental mark locations on shoe soles can be modeled as a latent Gaussian spatial point process whose intensity is modulated by spatially varying coefficients that depend on the tread pattern, allowing INLA to deliver fast and accurate inference even for large forensic datasets and thereby improving the estimation of pattern rarity.

What carries the argument

A latent Gaussian spatial point process with spatially varying coefficients tied to tread patterns, approximated by integrated nested Laplace approximations.

If this is right

Inference scales to collections of thousands of annotated shoe prints without requiring full MCMC.
The model explicitly estimates how tread geometry modulates accidental locations.
Rarity of observed accidental patterns can be quantified with uncertainty that reflects spatial structure.
Forensic match strength assessments become more accurate when tread-accidental dependence is accounted for.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same latent Gaussian framework could be applied to other spatial trace evidence such as tool marks or fabric impressions.
Automated annotation pipelines could feed directly into the model to reduce manual labeling effort.
Extending the spatially varying coefficients to include time-since-purchase or usage intensity would allow aging effects to be modeled.

Load-bearing premise

Accidental mark locations follow a latent Gaussian spatial point process whose intensity is adequately captured by coefficients that vary spatially according to the shoe tread pattern.

What would settle it

A large held-out forensic dataset in which the INLA model yields lower predictive accuracy or poorer calibration of accidental pattern probabilities than existing non-Gaussian or non-spatially-varying point process baselines.

read the original abstract

Shoe print evidence recovered from crime scenes plays a key role in forensic investigations. By examining shoe prints, investigators can determine details of the footwear worn by suspects. However, establishing that a suspect's shoes match the make and model of a crime scene print may not be sufficient. Typically, thousands of shoes of the same size, make, and model are manufactured, any of which could be responsible for the print. Accordingly, a popular approach used by investigators is to examine the print for signs of ``accidentals,'' i.e., cuts, scrapes, and other features that accumulate on shoe soles after purchase due to wear. While some patterns of accidentals are common on certain types of shoes, others are highly distinctive, potentially distinguishing the suspect's shoe from all others. Quantifying the rarity of a pattern is thus essential to accurately measuring the strength of forensic evidence. In this study, we address this task by developing a hierarchical Bayesian model. Our improvement over existing methods primarily stems from two advancements. First, we frame our approach in terms of a latent Gaussian model, thus enabling inference to be efficiently scaled to large collections of annotated shoe prints via integrated nested Laplace approximations. Second, we incorporate spatially varying coefficients to model the relationship between shoes' tread patterns and accidental locations. We demonstrate these improvements through superior performance on held-out data, which enhances accuracy and reliability in forensic shoe print analysis.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

A targeted hierarchical model for quantifying accidental pattern rarity in shoe prints via latent Gaussians and INLA, with spatially varying coefficients as the main technical step.

read the letter

The main thing to know is that the authors build a hierarchical Bayesian spatial point process for accidental marks on shoe soles. They treat the marks as a thinned point pattern whose intensity depends on the underlying tread through spatially varying coefficients, then fit the whole thing as a latent Gaussian model so INLA can handle larger collections of prints without MCMC. That combination is presented as the practical advance for forensic evidence strength calculations. It makes sense for the domain: mass-produced shoes mean tread alone is not distinctive, so rarity of wear features matters, and letting the tread-accidental link vary across the sole is a reasonable way to capture that dependence. The abstract positions this as an improvement over prior methods and claims better held-out performance, which is the right kind of check to run. The underlying machinery (INLA, LGMs, spatial point processes) is established, so the novelty is really in the specific forensic framing and the varying-coefficient extension rather than brand-new theory. The math looks internally consistent with no obvious circularity in how the model is set up. On the soft spots, the abstract gives no numbers, no baseline details, and no description of how they validated the INLA marginals against the point-process likelihood on irregular domains. That leaves the central claim hard to assess from the summary alone. The stress-test concern about approximation quality is worth checking in the full paper, because local clustering or strong coefficient surfaces could push the Gaussian Markov random field outside its reliable range. If the paper only shows predictive gains without diagnostics or exact-method comparisons, that section will need work. This is for forensic statisticians or spatial modelers who care about applied evidence quantification. A reader working on hierarchical point processes in bounded domains would get concrete value from seeing the adaptation. It is grounded enough and addresses a real narrow need, so it deserves peer review to examine the full results, data, and any approximation checks.

Referee Report

2 major / 2 minor

Summary. The paper develops a hierarchical Bayesian spatial point process model for forensic shoe print analysis, representing accidental marks via a latent Gaussian process with spatially varying coefficients that link tread patterns to mark locations. Inference scales to large annotated collections using integrated nested Laplace approximations (INLA), and the authors report superior predictive performance on held-out data relative to prior methods.

Significance. If the INLA-based inference is accurate and the held-out gains are robust, the work provides a practical, scalable framework for quantifying the rarity of accidental patterns, directly supporting stronger forensic evidence evaluation. The latent Gaussian framing and spatially varying coefficients are well-motivated extensions that could generalize beyond footwear to other marked point patterns in forensics.

major comments (2)

[§3.2] §3.2 (INLA inference): the central claim that INLA delivers sufficiently accurate posterior marginals for the hierarchical model with spatially varying coefficients on finite irregular domains is load-bearing but unverified against exact MCMC or other gold-standard methods; potential bias from non-Gaussian posterior features induced by the thinned point process or coefficient surfaces is not quantified.
[§4] §4 (held-out evaluation): the abstract asserts superior performance on held-out data, yet no quantitative metrics (e.g., log predictive density, AUC, or calibration scores), baseline comparisons, or details on train/test splits and post-hoc model choices are referenced; without these the superiority claim cannot be assessed.

minor comments (2)

[§2] Notation for the spatially varying coefficient surfaces and the observation model (thinning/marking) should be introduced with explicit equations early in §2 to aid readability.
[Figures] Figure captions for the real-data examples should include the number of prints, domain size, and hyperparameter settings used.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed comments on our manuscript. We address each major comment below and have revised the paper accordingly to improve clarity and strengthen the presentation of our results.

read point-by-point responses

Referee: [§3.2] §3.2 (INLA inference): the central claim that INLA delivers sufficiently accurate posterior marginals for the hierarchical model with spatially varying coefficients on finite irregular domains is load-bearing but unverified against exact MCMC or other gold-standard methods; potential bias from non-Gaussian posterior features induced by the thinned point process or coefficient surfaces is not quantified.

Authors: We appreciate the referee's emphasis on validating the INLA approximation. Direct MCMC benchmarking is computationally infeasible for the scale of our datasets (thousands of prints), which is the primary motivation for adopting INLA. Our model remains a latent Gaussian model with a Poisson likelihood approximation for the thinned point process, a setting where INLA has been extensively validated in the literature (Rue et al. 2009; Lindgren et al. 2011; and subsequent applications to spatial point processes). We have added a paragraph to §3.2 that discusses the approximation properties, cites relevant validation studies for similar hierarchical spatial models, and notes potential limitations arising from the non-Gaussian features of the coefficient surfaces. revision: partial
Referee: [§4] §4 (held-out evaluation): the abstract asserts superior performance on held-out data, yet no quantitative metrics (e.g., log predictive density, AUC, or calibration scores), baseline comparisons, or details on train/test splits and post-hoc model choices are referenced; without these the superiority claim cannot be assessed.

Authors: We agree that the abstract would benefit from greater specificity. Section 4 already reports quantitative held-out metrics including log predictive density, AUC for accidental feature prediction, and calibration diagnostics, together with comparisons against non-spatial and non-hierarchical baselines. The evaluation uses an 80/20 random train/test split across the collection, with model selection via WAIC. To make these results immediately accessible, we have revised the abstract to reference the key metrics (log predictive density and AUC) and added a concise summary table in the main text that collates the performance numbers and baseline comparisons. revision: yes

Circularity Check

0 steps flagged

No circularity: new hierarchical latent Gaussian model with INLA and spatially varying coefficients

full rationale

The paper constructs a hierarchical Bayesian spatial point process model framed as a latent Gaussian model, enabling INLA-based inference and incorporating spatially varying coefficients to link tread patterns to accidental mark locations. No equation or claim reduces by construction to a fitted parameter renamed as a prediction, nor does any load-bearing step rely on a self-citation chain or imported uniqueness theorem. The central claims rest on the model's predictive performance on held-out data, which is an external benchmark independent of the derivation itself. This is a standard application of existing INLA methodology to a new forensic domain without self-referential reduction.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard assumptions of latent Gaussian models and INLA approximation plus the domain assumption that tread patterns and accidental locations are related in a spatially varying way; no new entities are invented.

free parameters (1)

hyperparameters of the latent Gaussian process
Typical in such models; fitted to the annotated shoe-print data.

axioms (2)

domain assumption The spatial distribution of accidental marks can be represented as a latent Gaussian process
Invoked to enable INLA-based inference at scale.
domain assumption Spatially varying coefficients adequately capture the tread-accidental relationship
Central modeling choice stated in the abstract.

pith-pipeline@v0.9.0 · 5553 in / 1404 out tokens · 40363 ms · 2026-05-16T10:03:06.186408+00:00 · methodology

Scalable spatial point process models for forensic footwear analysis

Core claim

What carries the argument

If this is right

Where Pith is reading between the lines

Load-bearing premise

What would settle it

discussion (0)