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

TL;DR

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 is introduced.

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.

Figures (3)

• Figure 1: Visualized examples of well-aligned, weakly-aligned, and unaligned (raw data) scenarios in multispectral pedestrian detection. Unaligned pose a challenge as the detected persons in RGB (blue box) and in thermal (red box) does not overlap.
• Figure 2: Overall architecture of the proposed cross-modal semantic alignment fusion method.
• Figure 3: Visualized qualitative results: the same color means the same identity of a matched person across two different sensors.