Blobworld is a system for image retrieval based on finding coherent image regions that correspond to objects. Each image is automatically segmented into regions ("blobs") with associated color and texture descriptors. Querying is based on the attributes of one or two regions of interest, rather than a description of the entire image. To enable large-scale retrieval, blob descriptions are indexed using a tree. Because indexing in high-dimensional feature space is computationally expensive, a lower-rank approximation is used. Experiments show promising results for both querying and indexing. Image retrieval systems are evaluated based on their ability to find images based on objects, not just low-level features. Current systems often perform queries quickly but fail in precision and result understandability. Blobworld addresses this by segmenting images into regions that correspond to objects or parts of objects. The system automatically segments images without parameter tuning or manual region pruning. A complete online system for retrieving images from a collection of 10,000 Corel images is presented. Blobworld produces higher precision when querying for distinctive objects like tigers, zebras, and cheetahs compared to using color and texture histograms. Matching regions are highlighted, making false positives easier to understand. Query results are more interpretable and easier to refine with Blobworld. The speed of individual queries is also important. Blobworld features are indexed by projecting color feature vectors into lower dimensions. Queries using the index retrieve several hundred images and rank them using true distance, achieving results comparable to scanning the entire database. This paper reviews current image retrieval, outlines the segmentation algorithm, region descriptors, and querying system. It discusses indexing and presents experiments on querying and indexing performance. Related work includes systems based on low-level features, spatial relationships, and wavelet decompositions. Our indexing methods are based on those used in QBIC. Blobworld is distinct from color-layout matching as it is designed to find objects or parts of objects.Blobworld is a system for image retrieval based on finding coherent image regions that correspond to objects. Each image is automatically segmented into regions ("blobs") with associated color and texture descriptors. Querying is based on the attributes of one or two regions of interest, rather than a description of the entire image. To enable large-scale retrieval, blob descriptions are indexed using a tree. Because indexing in high-dimensional feature space is computationally expensive, a lower-rank approximation is used. Experiments show promising results for both querying and indexing. Image retrieval systems are evaluated based on their ability to find images based on objects, not just low-level features. Current systems often perform queries quickly but fail in precision and result understandability. Blobworld addresses this by segmenting images into regions that correspond to objects or parts of objects. The system automatically segments images without parameter tuning or manual region pruning. A complete online system for retrieving images from a collection of 10,000 Corel images is presented. Blobworld produces higher precision when querying for distinctive objects like tigers, zebras, and cheetahs compared to using color and texture histograms. Matching regions are highlighted, making false positives easier to understand. Query results are more interpretable and easier to refine with Blobworld. The speed of individual queries is also important. Blobworld features are indexed by projecting color feature vectors into lower dimensions. Queries using the index retrieve several hundred images and rank them using true distance, achieving results comparable to scanning the entire database. This paper reviews current image retrieval, outlines the segmentation algorithm, region descriptors, and querying system. It discusses indexing and presents experiments on querying and indexing performance. Related work includes systems based on low-level features, spatial relationships, and wavelet decompositions. Our indexing methods are based on those used in QBIC. Blobworld is distinct from color-layout matching as it is designed to find objects or parts of objects.