Revisiting Misalignment in Multispectral Pedestrian Detection: A Language-Driven Approach for Cross-modal Alignment Fusion

TL;DR

This paper presents a novel vision-language model-based framework for multispectral pedestrian detection that effectively handles heavy misalignment between modalities without complex calibration, improving real-world detection accuracy.

Contribution

It introduces a language-driven cross-modal alignment method using LVLMs to address misalignment issues in multispectral pedestrian detection.

Findings

Enhanced detection accuracy in heavily misaligned datasets

Elimination of traditional calibration pre-processing

Improved practical applicability of multispectral detection

Abstract

Multispectral pedestrian detection is a crucial component in various critical applications. However, a significant challenge arises due to the misalignment between these modalities, particularly under real-world conditions where data often appear heavily misaligned. Conventional methods developed on well-aligned or minimally misaligned datasets fail to address these discrepancies adequately. This paper introduces a new framework for multispectral pedestrian detection designed specifically to handle heavily misaligned datasets without the need for costly and complex traditional pre-processing calibration. By leveraging Large-scale Vision-Language Models (LVLM) for cross-modal semantic alignment, our approach seeks to enhance detection accuracy by aligning semantic information across the RGB and thermal domains. This method not only simplifies the operational requirements but also extends the practical usability of multispectral detection technologies in practical applications.