R cubed: Reconstruction, Raw, and Rain: Deraining Directly in the Bayer Domain
Learning directly on raw Bayer mosaics for rain deraining yields superior reconstructions compared to post-ISP processing. The study shows that ISP transformations are lossy and irreversible, making raw domain reconstruction better for preserving spatial and photometric information. This advocates an ISP-last paradigm for low-level vision and opens the way for end-to-end learnable camera pipelines.