Large scale metric learning from equivalence constraints

Abstract

In this paper, we raise important issues on scalability and the required degree of supervision of existing Mahalanobis metric learning methods. Often rather tedious optimization procedures are applied that become computationally intractable on a large scale. Further, if one considers the constantly growing amount of data it is often infeasible to specify fully supervised labels for all data points. Instead, it is easier to specify labels in form of equivalence constraints. We introduce a simple though effective strategy to learn a distance metric from equivalence constraints, based on a statistical inference perspective. In contrast to existing methods we do not rely on complex optimization problems requiring computationally expensive iterations. Hence, our method is orders of magnitudes faster than comparable methods. Results on a variety of challenging benchmarks with rather diverse nature demonstrate the power of our method. These include faces in unconstrained environments, matching before unseen object instances and person re-identification across spatially disjoint cameras. In the latter two benchmarks we clearly outperform the state-ofthe-art.

On-the-fly feature importance mining for person re-identification

“…Given the host of available appearance representations of colour, texture and shape, most existing distance metric learning based person re-identification methods take a GFI learning strategy [25,31,20,23,10,28,14]. Essentially, such techniques assume that certain features are universally more important in all circumstances, regardless of viewing condition changes between gallery and probe images and the specific visual appearance characteristics of a re-identification target person in the probe image.…”

“…In this study, we refer such universal feature weights selection schemes as learning top-down Generic Feature Importance (GFI). They can be learned either through boosting [10], rank learning [25,28], or distance metric learning [31,12,23,14].…”

A Study on Person Re-Identification System using Enhanced RNN

“…According to the results from table1, we completed the improved outputs while compared to efficient impostor-based metric learning [18] , Large scale metric learning [22] , Salient color names [19] and other approaches. For example, we reached much higher ranking rate comparing to KISSME method, this can be explained though the lower dimensional features and at the same time, same the calculating time.…”

“…The same person under the different camera have significant challenges such as viewpoint variation, illumination, deformation, occlusion, background clutter, which make the person re-identification a challenge problem. Figure 1 shows some challenging examples in a most used dataset PRID_450S [15] . Each red box contains the matched person from the dataset PRID_450S.…”

“…[24] show that the color as a single cue under extremely variable imaging conditions has relatively good performance in identifying persons, and present a novel illumination invariant feature representation which is based on the log chromaticity (log) color space. The existing features include: local binary patterns [25,11,12,20,14] , salient color names [26] , variations on color histograms [25,11,12,20,14,29] , Gabor features [14] , and local patches [28] .…”