Abstract
Existing multimodal image fusion (MMIF) datasets lack comprehensive coverage of adverse weather conditions. To address this limitation, we introduce AWMM-100K, a large-scale benchmark dataset constructed from RoadScene, MSRS, M3FD, and LLVIP, followed by controlled degradation processes to simulate rain, haze, and snow.
In addition, real-world data were captured using a DJI M30T drone equipped with high-resolution visible and thermal cameras. AWMM-100K contains over 187,699 images with weather conditions categorized into light, medium, and heavy intensities. The dataset supports research on multimodal image fusion as well as image restoration tasks such as dehazing, deraining, and desnowing.
AWMM-Text Update
We have updated the AWMM-Text annotations in AWMM-100K. The updated AWMM-Text release provides richer and more reliable textual supervision for adverse-weather multi-modality image fusion.
AWMM-Text provides paired textual annotations for adverse-weather multi-modality image pairs, including BLIP-generated global captions and ChatGPT-4-generated detailed descriptions. The global captions summarize scene semantics and degradation types, while the detailed descriptions provide fine-grained cues about objects, regions, and infrared-visible differences.
In total, AWMM-Text contains global and detailed textual descriptions for 8,500 multi-modality image pairs, together with their corresponding clean images. To ensure annotation quality, we manually screen the generated texts and inspect approximately 30% of them using clear pass/fail rules.
Dataset Overview
Image Fusion
Image Restoration
Real Scene
Compound Degradation
