Nonlinear spike-and-slab sparse coding for interpretable image encoding. PLOS ONE

2) Realistic occlusion dataset

This dataset contains realistic occlusions and controlled forms of sparse structure -- it is a realistic artificial dataset consisting of true occlusions (data is created by actual occlusions and not following any computational model for its generation). The data was generated using the Python Image Library (PIL) to draw hundreds of overlapping edges/strokes in a 256 × 256 pixel image: each stroke had an integer intensity between (1, 255), a width between (2, 4) pixels, and a length, starting, and ending position drawn independently from a uniform distribution. The image was then cut into overlapping D = 9 × 9 patches, each of which contained k ∈ (0, 5) overlapping strokes, for N = 61009. Gaussian noise of σ = 25 and μ = 0 was then independently added to each patch. Additionally, the dataset also contains the corresponding (automatically obtained) labels for each image, indicating the ground-truth number of occluding strokes k ∈ (0, 5) per image. Such a dataset represents and isolates challenging aspects of low-level image statistics that are present in all natural images. Particularly, it contains edges of varying intensities and their occlusions.
In the corresponding publication, data examples are shown in Figure 5.