Abstract:
The presence of haze significantly reduces the visibility of outdoor scenes, which not only
degrades the performance of many computer vision tasks but also compromises safety and
reliable decision-making in practical systems. In recent years, transformer-based models
have demonstrated remarkable performance in single-image dehazing by effectively capturing
long-range dependencies and global context. However, their effectiveness under real-world
haze conditions remains largely unexplored, as most models are trained and evaluated on
synthetic or artificially generated datasets. To address this limitation, we evaluate two
state-of-the-art transformer-based dehazing models, DehazeFormer and UMWTFormer, on
both real-world datasets (LM-Haze, RW-Haze and RTTS) and synthetic datasets (Dense-
Haze, RESIDE-OUT, O-HAZE and NH-HAZE), and analyse their performance using NIQE
and BRISQUE scores. Experimental results reveal a substantial performance drop, with
average dehazing effectiveness reduced by 40% on real-image datasets compared to synthetic
ones. This performance gap highlights a critical limitation in current dehazing models,
suggesting overfitting to synthetic data distributions and limited generalisation to real-world
complexities. These findings underscore the need for new models or training strategies explicitly
designed to cope with the variability and challenges inherent in real-world hazy images.
