MonoIR-RS: Infrared Remote Sensing Vision-Language Learning with CLIP and VLM Adaptation

TL;DR

The MonoIR-RS dataset and benchmark introduce infrared remote-sensing vision-language learning with CLIP and VLM adaptation. This approach improves infrared vision-language understanding by leveraging infrared-aware data construction and contrastive adaptation. The model achieves significant performance gains in infrared image understanding.

Why It Matters

The development of MonoIR-RS has practical significance in remote sensing applications, such as environmental monitoring and object detection. It enables more accurate analysis of infrared imagery, which can reveal valuable information not visible in RGB images. This can lead to improved decision-making in various fields, including ecology, geography, and emergency response.

Key Findings

• IR-aware adaptation improves CLIP recall by up to 12.8 points
• VLM captioning achieves 100% IR-cue coverage
• Synthesized infrared imagery outperforms grayscale conversion on the AVIID benchmark