Baseline Fusion for Image and Pattern Recognition—What Not to Do (and How to Do Better)

Abstract

Abstract:The ever-increasing demand for a reliable inference capable of handling unpredictable challenges of practical application in the real world has made research on information fusion of major importance; indeed, this challenge is pervasive in a whole range of image understanding tasks. In the development of the most common type-score-level fusion algorithms-it is virtually universally desirable to have as a reference starting point a simple and universally sound baseline benchmark which newly developed approaches can be compared to. One of the most pervasively used methods is that of weighted linear fusion. It has cemented itself as the default off-the-shelf baseline owing to its simplicity of implementation, interpretability, and surprisingly competitive performance across a widest range of application domains and information source types. In this paper I argue that despite this track record, weighted linear fusion is not a good baseline on the grounds that there is an equally simple and interpretable alternative-namely quadratic mean-based fusion-which is theoretically more principled and which is more successful in practice. I argue the former from first principles and demonstrate the latter using a series of experiments on a diverse set of fusion problems: classification using synthetically generated data, computer vision-based object recognition, arrhythmia detection, and fatality prediction in motor vehicle accidents. On all of the aforementioned problems and in all instances, the proposed fusion approach exhibits superior performance over linear fusion, often increasing class separation by several orders of magnitude.

Reimagining the central challenge of face recognition: Turning a problem into an advantage

“…These correspond to different exemplars f xy in Figure 3 and can be compared using the DFFS baseline. If, on the other hand, similarity is measured using the maximum correlation between subspace spans [32], the most similar modes of variation between two sets are readily extracted as the first pair of the canonical vectors between subspaces [33] and compared using the cosine similarity measure [34,35]. For manifold-to-manifold distances such as that of Lee et al [36] the most similar modes of variation are simply the nearest pairs of points on two manifolds, with the similarity of two points on the same manifold readily quantified by the geodesic distance between them.…”

“…As before, meta-feature training data is obtained using only face set pairs which are now represented by the corresponding covariance matrices. To extract training transitivity meta-features which correspond to same identity query-target comparisons, all reference set exemplars f qt iterate through and used to obtain f tq and f pq by anisotropically scaling them them using respectively the reference and proxy covariances (as in the original work [33]):…”