Spatial knowledge representation and retrieval in 3-D image databases

In multimedia retrieval applications such as architectural design, interior design, and real estate marketing, there exists a generic class of user queries that require retrieving images in the database that are spatially similar to the user query. We propose an image representation scheme (referred to as 3D spatial orientation graph or simply SOG) and an algorithm (referred to as SIM_3D) for retrieving 3D images of relevance (based on spatial similarity) to user queries from large image collections. Spatial similarity between a query and a database image is quantified based on the number as well as the extent to which the edges of SOG of the database image conform to the corresponding edges in the SOG of the query image. The time complexity of SIM_3D is Θ(|E_q|+|E_d|) where |E_q| and |E_d| are the number of edges in the query and database images. SIM_3D is robust in the sense that it can recognize translation and scale variants of apt image and these properties are shown formally. The effectiveness of SIM_3D is evaluated using a testbed image collection. The testbed comprises 60 images and are produced by generating 3 variants of each of the 15 original images. Image variants are produced by translation and scale transformations, and an arbitrary composition of these two transformations. The variants are designed to examine the robustness of the proposed algorithm. The results produced by the algorithm on a set of test queries are in agreement with the intuitively expected results