The PASCAL VOC Challenge datasets by Mark Everingham is a yearly dataset which has a central evaluation server and the final test data is not released. The latest edition (2012) contains 20 classes. The train/val data has 11530 images containing 27450 ROI annotated objects and 5034 segmentations. It is generally used for image classification, object detection, segmentation, action classification and person layout. There exists a development kit for evaluation. The dataset is large and very generic containing many different object categories with varying degree of difficulty (for example, car vs. chair). Only about 20% of the images have an overlap with other object instances or occlusions. The evaluation protocol is to detect all object categories and report the mean average precision, which is done by the development kit.