Dataset for "Spatial-Frequency Multi-Scale Sparse Attention for Single Image Raindrop Removal"

The dataset covers three public benchmark datasets: the AGAN-Data raindrop removal dataset (861 training image pairs, 58 testing image pairs), the SPAD rain streak removal dataset (14,000 training image pairs, 1,000 testing image pairs), and the Dense-Haze real-world haze removal dataset.

All images are cropped to 256×256 patches and organized into training and test sets following the original splits. Additionally, the package includes the complete training and testing code for the SFFormer model, pre-trained weight files, and ablation study configuration files, which can be used to reproduce all experimental results presented in the paper. This dataset is openly shared under the CC BY-NC-ND 4.0 license.