Content based image retrieval (CBIR) as it is known today has to deal with a number of challenges. Quickly summarized, the main challenges are firstly, to bridge the semantic gap between high-level concepts and low-level features using feedback, secondly to provide performance under adverse conditions. High-dimensional spaces, as well as a demanding machine learning task make the right way of indexing an important issue.
When indexing multimedia data, most groups opt for extraction of high-dimensional feature vectors from the data, followed by dimensionality reduction like PCA (Principal Components Analysis) or LSI (Latent Semantic Indexing). The resulting vectors are indexed using spatial indexing structures such as kd-trees or R-trees, for example.
Other projects, such as MARS and Viper propose the adaptation of text indexing techniques, notably the inverted file. Here, the Viper system is the most direct adaptation of text retrieval techniques to quantized vectors. However, while the Viper query engine provides decent performance together with impressive user-feedback behavior, as well as the possibility for easy integration of long-term learning algorithms, and support for potentially infinite feature vectors, there has been no comparison of vector-based methods and inverted-file-based methods under similar conditions.
In this publication, we compare a CBIR query engine that uses inverted files (Bothrops, a rewrite of the Viper query engine based on a relational database), and a CBIR query engine based on LSD (Local Split Decision) trees for spatial indexing using the same feature sets.
The Benchathlon initiative works on providing a set of images and ground truth for simulating image queries by example and corresponding user feedback. ·when performing the Benchathlon benchmark on a CBIR system (the System Under Test, SUT), a benchmarking harness connects over internet to the SUT, performing a number of queries using an agreed-upon protocol, the multimedia retrieval markup language (MRML). Using this bench-mark one can measure the quality of retrieval, as well as the overall (speed) performance of the benchmarked system.
Our Benchmarks will draw on the Benchathlon's work for documenting the retrieval performance of both inverted file-based and LSD tree based techniques. However, in addition to these results, we will present statistics, that can be obtained only inside the system under test. These statistics will include the number of complex mathematical operations, as well as the amount of data that has to be read from disk during operation of a query.