Person re-identification by probabilistic relative distance comparison

Abstract

Matching people across non-overlapping camera views, known as person re-identification, is challenging due to the lack of spatial and temporal constraints and large visual appearance changes caused by variations in view angle, lighting, background clutter and occlusion. To address these challenges, most previous approaches aim to extract visual features that are both distinctive and stable under appearance changes. However, most visual features and their combinations under realistic conditions are neither stable nor distinctive thus should not be used indiscriminately. In this paper, we propose to formulate person re-identification as a distance learning problem, which aims to learn the optimal distance that can maximises matching accuracy regardless the choice of representation. To that end, we introduce a novel Probabilistic Relative Distance Comparison (PRDC) model, which differs from most existing distance learning methods in that, rather than minimising intra-class variation whilst maximising intra-class variation, it aims to maximise the probability of a pair of true match having a smaller distance than that of a wrong match pair. This makes our model more tolerant to appearance changes and less susceptible to model over-fitting. Extensive experiments are carried out to demonstrate that 1) by formulating the person re-identification problem as a distance learning problem, notable improvement on matching accuracy can be obtained against conventional person re-identification techniques, which is particularly significant when the training sample size is small; and 2) our PRDC outperforms not only existing distance learning methods but also alternative learning methods based on boosting and learning to rank.

Multiple-Shot Person Re-Identification by Pairwise Multiple Instance Learning

“…To show the impact of pairwise learning scheme, we implemented a 'MIL-based' method which only replaced our learning model with MIL boosting proposed in [13]. PMIL obtains much better result than 'MIL-based' method, which is also consistent with the conclusion in [7], [8], [10] that the pairwise ranking is suitable in handling visual ambiguity.…”

“…3 (e)-(f), the most erroneous recognition of PMIL are due to significant color variations caused by different environments since we only employ color features. Incorporating other texture features, such as Gabor and Schmid filters [7], [8] may improve the performance. Visual ambiguity exists in this dataset due to a few people wearing black suits and drastic illumination changes (e.g.…”

“…This criteria shares the similar spirit with [7], [8], [10], [11], in terms of treating person reidentification as a ranking problem to tackle visual ambiguity. However, we integrate pairwise comparison into appearance learning while they focus on metric learning between two cameras.…”

Bi-level Relative Information Analysis for Multiple-Shot Person Re-Identification

“…In order to further validate the capability of SCNNM dissimilarity itself, we conduct experiments on its cooperation with several possible combinations of features and low-level set-to-set distances, such as "Ori SCNNM(MPD)", "Ori SCNNM(CHISD)", and "TPCR SCNNM(MPD)". Moreover, we demonstrate the capability of our method by comparing with typical state-of-the-art methods as well, including the unsupervised method MRCG, and the supervised methods MCC [16], RankSVM [17], RDC [18], and SBDR [4]. Experimental results are illustrated by the "Cumulative Matching Characteristic" (CMC) curve, which visualizes the expectation of the correct match at each rank based on the ranking of each of the corpus w.r.t.…”

Person Re-Identification by Spatial Pyramid Color Representation and Local Region Matching

“…The first kind is to optimize the ranking of a true match, generally for a given dataset. Ensemble RankSVM [1] and probabilistic relative distance comparison [2] were developed to address the scalability and over-fitting problems caused by insufficient training samples. However, this kind of method is not extensible in a camera network because the scene structure and viewpoint may vary greatly for each camera.…”