부산대학교 도서관

Utilizing the multi-modality of RGB and thermal images is valuable in various situations where visibility is reduced due to bad weather, fog, or lighting conditions. In previous semantic segmentation work, high recognition performance has been achieved by combining features extracted from RGB images with features from thermal images, but the performance is limited because the output features are combined without considering the model weights according to the characteristics of the modalities. In this paper, we configure the networks to have common weights during the learning process of RGB and Thermal images, and configure the modality features with more information to lead the other modalities in the loss calculation step. We also configure complementary masking modules and residual networks to achieve a performance improvement of 12.25 on the FMB dataset.