The SUN09 dataset consists of 12,000 annotated images covering more than 200 object categories. The dataset contains natural, indoor, and outdoor images. Each image contains an average of 7 different annotated objects, and the average area occupied by each object is 5 % of the image size. The frequency of object categories follows a power-law distribution.

The dataset contains two major benchmarks:

1. Used to evaluate the overall target recognition system, including:

• static_sun09_database: 12,000 annotated images
• static_sun_objects: additional images used for training the baseline detector (not used for training the context model)
• out_of_context: 42 out-of-context images

2. Contextual models for evaluating the output of a pre-computed baseline detector, which contains:

• The file name corresponds to [(test/train)/objectCategory/imageName.txt]
• Each line in the text file shows the bounding box position and score of a candidate window: [ x1 y1 x2 y2 score ]
• We use 4,367 training images and 4,317 testing images. Each group has the same number of images per scene category.
• Load sun09_detectorOutputs.mat for the baseline detector outputs and sun09_groundTruth.mat for the ground truth annotations.

This dataset was published by MIT at the 2010 IEEE CVPR.