Properties of optimally weighted data fusion in CBMIR

Content-Based Multimedia Information Retrieval (CBMIR) systems which leverage multiple retrieval experts (E_n) often employ a weighting scheme when combining expert results through data fusion. Typically however a query will comprise multiple query images (I_m) leading to potentially N × M weights to be assigned. Because of the large number of potential weights, existing approaches impose a hierarchy for data fusion, such as uniformly combining query image results from a single retrieval expert into a single list and then weighting the results of each expert. In this paper we will demonstrate that this approach is sub-optimal and leads to the poor state of CBMIR performance in benchmarking evaluations. We utilize an optimization method known as Coordinate Ascent to discover the optimal set of weights (|E_n| ⋅ |I_m|) which demonstrates a dramatic difference between known results and the theoretical maximum. We find that imposing common combinatorial hierarchies for data fusion will half the optimal performance that can be achieved. By examining the optimal weight sets at the topic level, we observe that approximately 15% of the weights (from set |E_n| ⋅ |I_m|) for any given query, are assigned 70%-82% of the total weight mass for that topic. Furthermore we discover that the ideal distribution of weights follows a log-normal distribution. We find that we can achieve up to 88% of the performance of fully optimized query using just these 15% of the weights. Our investigation was conducted on TRECVID evaluations 2003 to 2007 inclusive and ImageCLEFPhoto 2007, totalling 181 search topics optimized over a combined collection size of 661,213 images and 1,594 topic images.