SegIR: Semantic-Aware Infrared Image Generation

Translating RGB images into their infrared (IR) counterparts is a crucial task due to the broad utility of IR imagery in applications such as object detection and tracking. However, existing deep learning methods often overlook the distinct semantic and thermal properties of the IR domain. In this work, we present SegIR, a generative adversarial network-based (GAN) model that leverages external semantic segmentation information to generate IR images while preserving the thermal characteristics of the IR spectrum. Our proposed method outperforms existing GAN and transformer-based baseline models across multiple evaluation metrics on FLIR and KAIST datasets. Furthermore, we propose a novel variant of SegIR that synthesizes IR images directly from segmentation maps alone. To the best of our knowledge, this is the first approach to enable IR image generation from purely semantic input in the context of IR translation. Datasets and codes are available at: https://github.com/bleuuuuuuu/SegIR.