Multi-Source Visual Language Model Fusion for Indoor Geolocation Reliability

Νικόλαος Κωνσταντίνου

doi:10.1109/ACCESS.2026.3656619

2026-01-21

Νικόλαος Κωνσταντίνου

Theodoros Semertzidis

Petros Daras

DOI：10.1109/ACCESS.2026.3656619

原文链接

原文求助

摘要

En 中文

Black-box indoor geolocation systems lack interpretable mechanisms to assess prediction reliability, hindering deployment in forensic investigations, legal proceedings, and intelligence operations where users must distinguish trustworthy from unreliable outputs. This work presents a two-stage investigation with empirically-derived architectural validation infrastructure addressing this gap. The first stage evaluates whether Visual Language Models (VLMs) can provide interpretable reliability assessment through architectural reasoning. Applied to 2,397 residential indoor images from verified Airbnb listings across six countries, the analysis demonstrates that VLM confidence creates meaningful discrimination within black-box predictions (24-point gap in accuracy between high- and low-confidence subsets). However, this discrimination exhibits geographic variability (ranging from −4.6 to + 50.5 percentage points across countries), limiting standalone deployment reliability. The second stage addresses these limitations through learned multi-source fusion that combines calibrated geographic probabilities, VLM confidence, and architectural segmentation to detect consensus patterns predictive of reliability. This approach achieves 82.2% accuracy at 5% coverage and sustains over 80% precision at 10% coverage, substantially outperforming single-source baselines. Ablation experiments confirm that calibration, interaction terms, and the fusion architecture each contribute significantly to performance, demonstrating that robust reliability assessment for indoor geolocation emerges from calibrated multi-source consensus rather than standalone confidence scores. Code will be made available.

Keyword:

Confidence calibration

indoor geolocation

interpretability

multi-source fusion

reliability assessment

vision language models