Abstract
RGB-Thermal (RGBT) semantic segmentation is an emerging technology for identifying objects and materials in high dynamic range scenes. Thermal imaging particularly enhances feature extraction at close range for applications such as textile damage detection. In this paper, we present RGBT-Textile, a novel dataset specifically developed for close-range textile and damage segmentation. We meticulously designed the data collection protocol, software tools, and labeling process in collaboration with textile scientists. Additionally, we introduce ThermoFreq, a novel thermal frequency normalization method that reduces temperature noise effects in segmentation tasks. We evaluate our dataset alongside six existing RGBT datasets using state-of-the-art (SOTA) models. Experimental results demonstrate the superior performance of the SOTA models with ThermoFreq, highlighting its effectiveness in addressing noise challenges inherent in RGBT semantic segmentation across diverse environmental conditions. We make our dataset publicly accessible to foster further research and collaborations.