Trajectory-Anchor Optimization for Overconfident Thermal Visual Place Recognition: Zero-Leakage OOD Auditing and Kidnapped-Robot Recovery

Jul 16, 2026

Upload by Aria

Modern thermal visual place recognition (TIR-VPR) frontends based on foundation-model descriptors deliver strong closed-set retrieval, yet fail under out-of-distribution or unmapped conditions due to an overconfident forced-matching mode that yields locally consistent but false loop candidates. Classical multi-hypothesis tracking can reduce spatial ambiguity but introduces prohibitive real-time latency. Trajectory-Anchor Optimization (TAO) replaces MHT with batched SE(2) Procrustes alignment solved via tensor-vectorized, single-invocation batched SVD, enabling strictly bounded per-frame computation. Passive geometric consistency further acts as a conservative rejector. Zero-leakage auditing shows a micro-scale separation limit near 5m, while a macroscopic convergence basin around 10m reliably suppresses catastrophic topological breaks.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages11

Reading Time28 min

Introduction

Overconfident forced-matching failure mode in thermal VPR

Limits of multi-hypothesis tracking for real-time robotics

Trajectory-Anchor Optimization (TAO) as a bounded multi-view backend

Zero-leakage auditing and operational boundary

Micro-scale indistinguishability near 5m

Macroscopic convergence basin around 10m

Filtering false accepts to prevent graph corruption

arXiv :2607 .04745v 1 [ cs .RO] 6 Jul 2026 Trajectory-Anchor Optimization for Overconfident Thermal Visual Place Recognition: Zero-Leakage OOD Auditing and Kidnapped-Robot Recovery Zhiyuan Lu Kanji Tanaka July 7, 2026 Abstract Modern thermal visual place recognition (TIR-VPR) frontends, particularly foundation-model-based descriptors such as AnyThermal-SALAD, exhibit remarkable retrieval performance as closed-set candidate generators. However, their structural vulnerability lies in an overconfident forced-matching failure mode: under out-ofdistribution (OOD) or unmapped conditions, these deep models generate highly plausible, locally consistent yet false loop candidates without a noticeable drop in similarity scores. While classical multi-hypothesis tracking (MHT) backends can mitigate such spatial ambiguities by maintaining divergent trajectory beliefs, their exponential hypothesis management overhead introduces prohibitive computational latency, fundamentally violating real-time robotic constraints. To bridge this gap, we present Trajectory-Anchor Optimization (TAO) . Crucially, to counter the combinatorial challenge of evaluating a large pool of parallel hypotheses (e.g. , K = 100), TAO compresses the multi-view temporal verification into a batched SE(2) Procrustes alignment problem. By leveraging tensor-level vectorization and single-invocation batched Singular Value Decomposition (SVD), this formulation entirely bypasses the dynamic tree expansion and memory allocation overhead inherent to MHT, guaranteeing a strictly bounded per-frame execution loop of O (KN) while maintaining rigid multi-view geometric consistency. While this passive temporal verification effectively bounds the search space, we show that incorporating rigid multi-view geometric consistency establishes a risk-averse, conservative rejecting layer capable of filtering unstructured spatial hallucinations without requiring active exploration strategies. Under a strict zero-leakage evaluation protocol, we demonstrate that while a purely passive geometric backend cannot mathematically separate metric localization errors from highly coherent hallucinations at a micro-scale (5m) due to local visual ambiguities, TAO serves as a highly efficient, lightweight fail-safe filter at a macro-scale (large map-out scenarios) . The visual foundation model’s overconfident hallucinations within a tight 5m radius often possess a locally consistent geometry that deceives rigid multi-view alignment. However, beyond this micro-scale threshold, the K = 100 disparate retrieval hypotheses disperse spatially across the global map. This spatial dispersion breaks the rigid temporal co-visibility constraint within the sliding window (N = 20), causing the joint optimization residual to escalate sharply. Consequently, while the passive backend exhibits an information-theoretic limit atthe sub-5m resolution, it establishes a distinct macroscopic convergence basin (10m) where multi-view geometric consistency reliably isolates catastrophic topological breaks and suppresses critical false acceptances, defining a transparent operational boundary for preventing graph corruption. 1 Introduction Thermal infrared (TIR) cameras are highly attractive for long-term autonomous navigation because they remain informative across severe diurnal illumination changes. Driven by deep foundation models, recent TIR visual place recognition (VPR) frontends, such as AnyThermal-SALAD [11], have substantially improved closed-set retrieval accuracy. Yet, an architectural paradox remains: a stronger retrieval frontend does not automatically guarantee a safer SLAM backend. Due to the low-texture nature of thermal imagery, these models suffer from a critical flaw of overconfident forced matching—consistently yielding false loop candidates backed by falsely high similarity scores that exhibit a deceptive, smooth continuity even when traversing completely unmapped out-of-distribution (OOD) environments.

FAQs

What problem does TAO address in thermal visual place recognition?

TAO targets overconfident forced-matching failures in thermal VPR, where models generate false loop candidates that remain locally consistent and may not show a similarity-score drop under OOD or unmapped conditions.

Why are traditional multi-hypothesis tracking backends considered unsuitable for real-time deployment?

Multi-hypothesis tracking can require an exploding hypothesis tree during extended unmapped periods, creating intensive computation and collapsing real-time performance.

How does TAO achieve bounded computation while verifying multiple hypotheses?

TAO evaluates K parallel trajectories simultaneously by converting temporal verification into batched SE(2) Procrustes alignment problems, solved with tensor-level vectorization and a single batched SVD, avoiding dynamic tree expansion and memory overhead.

Jul 16, 2026

Upload by Aria