Filled, dilated, and textured MPEG-7 datasets

filled-dilatedMPEG-7: A minor augmentation of MPEG-7 dataset to improve mask contiguity. 

Binary dilation of each MPEG-7 mask by 3 pixels, topologically simplified by binary-infilling holes. The external boundary of the shape was extracted as a contour (x,y coordinates) using marching squares (scikit-image find_contours). For a minority of shapes, the dilation and binary-infilling still produces multiple disconnected masks (due to a fragmented contour). In these instances, we recommend only using the largest connected mask.

filled-dilatedMPEG-7_x_normalizedBrodatz Dataset: Combination of selected padded-filledMPEG-7 shape classes and normalized Brodatz features. 

Five shape classes from MPEG-7 were chosen by equally sampling the measured eccentricity, which was used as a proxy measure of shape complexity. Five normalized Brodatz texture images were chosen randomly. To texture the MPEG-7 shapes, their images were all standardized by rescaling to 128x128 pixels. Brodatz texture images were rescaled to 192x192 images. Then for each of the 20 unique images in each MPEG-7 class, we take each 192x192 Brodatz texture image, sample 20 random 128x128 cropped patches and multiply with the MPEG-7 image to create 20 different textured variants of the same basic texture and shape. This gives a total dataset of 10,000 image patches (5 shape classes x 20 images per shape class x 5 texture classes x 20 per texture class).