The goal of this project is to create image retrieval systems based on the objects that appear in the images. Offline, we learn a probabilistic model for the general appearance of the database domain (i.e., the appearance of typical images in the database). Then, when the user presents some examples of an object's appearance, we learn a probabilistic model for that object and search the database for images containing that object. Each image is searched using a windowed scan over location and scale. Images are returned to the user ranked by the posterior probability of the object class. The user can the provide feedback on the search results, allowing the system to further refine its search. This work is performed by Derek Hoiem, Rahul Sukthankar (Intel), Henry Schneiderman (CMU), and Larry Huston (Intel). The work is supported by Intel Research Pittsburgh.
List of Images Used for Experiments
We tested our system using ten categories from the Corel image database.
We removed images that did not contain the complete object of interest at sufficient resolution for that category.
For the windsurfing category, we also used images from sailboarding so that there was a sufficient number of images for testing.
List of Blobs Used for Blobworld Comparison
We compared our system to Blobworld using the available online code for Blobworld. To test, we needed to label the "blob" containing the object of interest for each image file. These are the blobs used for each image when running the normal segmentation process performed by the Blobworld segmentation code.
D. Hoiem, R. Sukthankar, H. Schneiderman, and L. Huston, "Object-Based Image Retrieval Using the Statistical Structure of Images", IEEE Conference on Computer Vision and Pattern Recognition, 2004.
pdf