The MSRC (Microsoft Research Cambridge) Object Recognition Image Dataset available here, research.microsoft.com/en-us/project/ObjectClassRecognition/ contains pixel-wise labeled images (591 images, 23 object classes). For our evaluation, we select the major topics contained in the images and refine the ground truth for that object. For example, a sheep image contains sheep, grass, sky, etc.; however, we only evalute the quality of the sheep polygons.
Here are the visualizations of their data using Markup SVG.
OSD = overlap score difference between the given ground truth and our computed ground truth.
sheepOSD=0.889 (30 images)
birdOSD=0.832 (34 images)
carOSD=0.907 (30 images)
cowOSD=0.860 (30 images)
treeOSD=0.908 (30 images)