Strange97/AODLib-EAOD

OmniLens releases the AODLib-EAOD dataset covering diverse optical degradation patterns from different lens types, enabling models that generalize impressively across varied aberration distributions. We anticipate further work on universal blind aberration correction model designs building on this dataset. The AODLib-EAOD data contains 3 parts. Here are the descriptions:

• -Imgs: This training ready dataset comprises 2,111 pairs of degraded and ground truth images. Each pair is produced by randomly sampling one lens from the 1,000 sampled AODLib-EAOD lenses and simulating aberration, distortion, and ISP. All pre-trained models reported in the paper are trained on this paired set.
• -Lens_Descriptions: It contains PSFs (-psf) computed at 64 uniformly normalized fields of view for the 1,000 selected lenses, together with relative illumination distributions (-ill) and distortion maps (-distort), which can be used directly to simulate aberration degradation and distortion for each lens.
• -All_Lenses: The unsampled raw lens library. It likewise includes the PSFs, relative illumination distributions, and distortion maps that can be used directly for simulation. The RMS spot radius and related specifications for each lens are annotated in the filenames. We recommend that interested researchers use these data to filter desired aberration degradations and construct datasets.

Illustration of '-psf' data

Type: Torch.tensor

Shape: [64, 3, H, W];

Description: PSFs at 64 normalized fields of view from center to edge, with 3 channels (RGB) and size H×W. The provided PSF orientation points from the image center toward the downward direction, and you need to rotate the PSFs according to the specific FoV layout to complete the simulation.