You are here
Ground-Truth for PonceDB - labelled images for object recognition tests
To contribute in testing of vision-based object recognition systems, TeV makes available for download masks and labels of an image database, named here Ponce's Object Recognition dataset, we have added for testing the performance of our object recognition system, MEMORI.
Ponce's database consists of 161 images depicting 8 objects in different poses against an almost uniform background and 51 test cluttered images containing rescaled, rotated, partially occluded and differently illuminated instances of the objects. The images are color JPEGs, the resolutions are 1.2 Mpix (1280x960) and 3.7 Mpix (2200x1700). The Ponce's Object Recognition Database is freely available starting from Object Recognition Database - Robotics and Computer Vision Laboratory Beckman Institute, University of Illinois at Urban-Champain, USA.
The 8 classes of objects are: 1. apple; 2. bear; 3. rubble; 4. salt; 5. shoe; 6. spider-man; 7. truck; 8. vase.
Reference: Fred Rothganger, Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. 3D Object Modeling and Recognition Using Local Affine-Invariant Image Descriptors and Multi-View Spatial Constraints. International Journal of Computer Vision, 66(3):231-259, March 2006
To evaluate the performance of our object recognition systems, we have tagged each Ponce's image (rescaled by a factor 0.5) with ground-truth information, with a labelled map, where each labelled segment corresponds to one object or to background. We made therefore available:
- Masks of the objects - The object views have been separated from the background by a semi-automatic method (thresholding strategy and manually refinement).
- Ground-truth - To each object class in the database an label has been assigned: 1 for apple, 2 for bear, ..., 8 for vase. For each of the 51 test images the objects (or the parts of them in case of occlusions) have been manually separated from the background. For each image an label map has been built: the background corresponds to the region with value 0, whereas the segments correspondent to an object have an integer value greater than 0.
- An annotation file (LabelInfoPonceTest.dat) which contains the labels of the objects depicted in each test image. More precisely, each row of this file is related to one of the test image and contains:
- the identifier I of the test image (01, 02,..., 51);
- the number of objects contained in the image;
- the labels of the objects contained in the image; in particular, the label J identifies the object that occupies the region with value J in the labelled map.
Therefore, the GroundTruth-for-PonceDB archive comprises five parts: reduced images and respective labelled maps, both for database and test images, and an annotation file for the test image archive.
To download our GroundTruth-for-PonceDB please follow this link.
To download the images of the Ponce's DB object database (objects and tests) use the following link: Object/Test images archive (.tgz format - 108 MB)
Remark: this data can only be used for research or academic purposes. For a complete test on recognition and pose estimation of objects in images you need a rescaled version (factor 0.5) of the Ponce's images.