Synthetic Edges for Corner Training DATA

The training dataset is composed by 31x31pixels images of synthetic edges of 38160 (73.6%) CORNERS and 13680 (26.4%) NO-CORNERS. We generated, from the central pixel, two line segments, L1 and L2, that are oriented by the angles a1 and a2 respectively as shown in "line-segments.png". The angles vary from 0° to 355°, in increments of 5°. When the inner angle between L1 and L2 is less than 140°, the descriptor that is associated to the central pixel belongs to the CORNER Class. Otherwise, a greater angle is considered a straight line and the central pixel belongs to the NO-CORNER class. Aiming to obtain a noise-robust corner detector, we generated ten images with salt and pepper noise with 0.04 as noise density, for each combination (L1,L2) as shown in "dataset.png".