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

Forum topic · 小凯 · 2026-07-09

Summary

MonoIR-RS is a large-scale infrared remote-sensing vision-language dataset and benchmark introduced by Jiaju Han, Ma Yaqi, and Yahui Chai (arXiv 2507.06827, July 2025). Infrared imagery captures intensity structure, object-background contrast, and illumination-invariant cues that are often invisible in RGB, yet most remote-sensing vision-language resources focus on visible-band semantics. Built from the same source pool and split as FusionRS, MonoIR-RS keeps the infrared image as the model-facing modality, producing 600,000 synthesized infrared images and 59,032 retained IR-aware caption records whose captions emphasize grayscale structure and IR-style contrast rather than RGB appearance. The authors show the synthesized infrared images match real thermal imagery more closely than grayscale conversion on the AVIID benchmark. Fine-tuning five CLIP backbones and six VLM backbones, IR-aware adaptation improves CLIP average recall by up to 12.8 points and drives infrared-cue coverage in VLM captions to 100% while reducing residual RGB color leakage to near zero, providing a controlled testbed for aligning infrared evidence with language.

Paper Overview

Field: Computer Vision
Authors: Jiaju Han, Ma Yaqi, Yahui Chai
Released: 2025-07-09
arXiv: 2507.06827

Abstract

Infrared remote-sensing imagery captures intensity structure, object-background contrast, and illumination-invariant cues often invisible in RGB imagery. Yet, most remote-sensing vision-language resources and models focus on visible-band semantics, leaving infrared vision-language understanding underexplored. MonoIR-RS is a large-scale infrared remote-sensing vision-language dataset and benchmark that couples IR-aware data construction with CLIP-style contrastive adaptation and VLM instruction tuning.

Key Details

Built from the same source pool and split as FusionRS, keeping the infrared image as the model-facing modality.
Yields 600,000 synthesized infrared images and 59,032 retained IR-aware caption records; model experiments use this retained language-supervision subset, with captions rewritten around grayscale structure and infrared-style contrast rather than RGB appearance.
Synthesized infrared images are shown to be substantially closer to real thermal imagery than grayscale conversion on the AVIID benchmark.
The authors fine-tune five CLIP backbones and six VLM backbones, calibrating against zero-shot behavior:
IR-aware adaptation improves CLIP average recall by up to 12.8 points.
Infrared-cue coverage in VLM captions is driven to 100%, while residual RGB color leakage drops to near zero.

By isolating the infrared modality from RGB-IR multimodal learning, MonoIR-RS provides a controlled, reproducible testbed for aligning infrared remote-sensing evidence with language.

Links: arXiv page