INRIA Annotations for Graz-02 (IG02)
INRIA Annotations for Graz-02 (IG02) is a reedition of the popular natural-scene object category dataset prepared at Graz University of Technology. The new annotations created at INRIA are aimed to be object-oriented and more precise. This dataset includes photos with objects of high complexity and high intra-class variability on cluttered backgrounds.
The images were collected by Andreas Opelt and Axel Pinz. The annotation was led by Marcin Marszałek and Cordelia Schmid. We thank Axel Pinz for his support and comments.
The original images have the resolution of 640x480 or 480x640 pixels. Red color marks the visible object parts and green is used for occlusions.
Our team re-annotated cars, bicycles and people on the original set of images. Only some portrait images incorrectly saved as landscape images were corrected. For each object a segmentation mask was drawn, which includes occluded object parts (marked with a different color). Each object was marked "truncated" when it was cut by the image edge, "multiple" when it could not be separated from other objects of this class and "difficult" if it was hard to notice or segment. The images were then considered as suitable for training (when there was at least one non-truncated and non-multiple object in the image) and testing (if all the objects in the image could be individually segmented). As those lists had some overlap, we have randomly partitioned it to create a suggested balanced split into a training set and a test set.
Download & statistics
To cite the dataset please use:
Marcin Marszałek and Cordelia Schmid. Accurate Object Localization with Shape Masks. IEEE Conference on Computer Vision & Pattern Recognition, 2007.
A. Opelt, A. Pinz, M. Fussenegger, P. Auer. Generic Object Recognition with Boosting. IEEE Transactions on Pattern Recognition and Machine Intelligence (PAMI), Vol.28, No. 3, March 2006.
Please let us know about publications in which you use our dataset, so we can refer to them here.
The tool we used to annotate images was created by Alexander Kläser and it is available (source code included) at his website. It was designed for quick and accurate drawing of object segmentations.
Bugs and updates
If you spot any errors in the annotations, please report them to Marcin Marszałek. Corrections of the segmentations are especially welcome. When the number of corrections becomes significant, we will release an update.