BSDS500 Contour Detection and Semantic Segmentation Dataset

BSDS 500 is a dataset for contour detection and semantic segmentation research, containing 12,000 manually labeled 1,000 Corel dataset images collected from 30 human subjects.

Half of the segmentations were obtained by presenting color images to the subject; the other half came from presenting grayscale images. A public benchmark based on this data consists of all grayscale and color segmentations of 300 images, split into 200 training images and 100 test images.

This dataset was released by the University of California, Berkeley in 2001.

Related papers: "A Database of Human Segmented Natural Images and its Application to Evaluating Segmentation Algorithms and Measuring Ecological Statistics"