Fig 5. Synthetic occlusion dataset and cut–out original and noisy patches.

Examples taken from the occlusion dataset. A shows an original noise-free image of generated occluding strokes of random width, pixel intensity, and starting/ending points. B shows a handfull of overlapping image patches cut from the original, noise-free data. C shows examples of the noisy training data, with independent σ = 25 noise added to B.