One major challenge is lack of paired RGB-IR data.
F-ViTA (Foundation Model Guided Visible-to-Infrared Translation)
F-ViTA generates thermal (IR) images from RGB inputs using object-level and scene-level conditioning guided by foundation models. This allows models to learn realistic heat patterns and improves performance in settings where paired RGB-IR data is limited.
TherA (Thermal-Aware Visual-Language Prompting)
TherA enhances RGB-to-thermal translation by incorporating temperature-aware prompts, enabling more controllable and semantically meaningful thermal image generation. It improves how well generated IR images reflect real-world thermal cues, making them more useful for downstream tasks.
RARL (Reasoning-Aware Reinforcement Learning with LoRA)
RARL improves vision-language models by encouraging structured reasoning through reinforcement learning with multiple reward signals, including accuracy and reasoning quality. Combined with LoRA for efficient fine-tuning, it enables better generalization and transparent decision-making under limited data and compute constraints.