MPEG-7 Homogeneous Texture Descriptor Demo

Demonstrates content-based similarity retrieval using the MPEG-7 Homogeneous Texture Descriptor. The dataset consists of 54 large aerial photographs of the Santa Barbara, California region. Each grey-scale image is divided into 128x128 pixel non-overlapping tiles for a total of 90,774 tiles. A 60 dimensional texture feature vector is extracted from each tile. These feature vectors are used to perform similarity retrieval in a query-by-example framework.

This demo illustrates two points. First, that the MPEG-7 descriptor successfully characterizes homogeneously textured regions. Second, that high-resolution remote sensed imagery contains a wide selection of regions with characteristic texture signatures. Examples include agricultural fields, housing developments, transportation components, etc.

Click here to try it.

Category-Based Image Search on the WWW

Demonstrates image search on the WWW based on three sources of information: 1) associated text such as HTML, 2) image content primitives such as color and texture, and 3) a readily available hierarchical category structure of the WWW. All of this information is automatically "spidered" and indexed--no manual annotation or guidance is necessary. A relevance feedback loop is used to automatically refine the query as well as learn the relative weighting of the image content primitives.

The demo contains over 600,000 images collected from the WWW. We focused mainly on shopping/clothing categories so some queries to try include hats, shoes, etc.

This demo is an example of how different data modalities can be combined to help overcome their individual limitations. In particular, the attendant text on the webpages can be used to address the semantic gap faced by content-based image search.

Click here to try it.