There are quite a few datasets for training and testing computer vision algorithms, such as the ImageNet database and PASCAL databases. Most of the these databases are created from images with various restrictive licenses. Are there databases (with training data, such as classifications, segmentations, etc.) that are based on open access images (ie. Creative Commons, public domain, or the like) or are there meta-projects to help create such databases?
A possible example would be a project to collect open access images that are well suited for computer vision algorithms (particularly natural everyday images, not mathematical plots, cartoons, etc). I'm mainly interested in data for tasks that humans can perform, such as object detection, classification, object segmentation, etc.
mldata.org is perhaps the closest thing that I've found. They give the licenses of the datasets, although they don't provide a way to filter by license or type of license and it doesn't seem like they are always accurate.