Transitive Re-identification

Person re-identification (ReID) problem has lately received increasing attention especially due to its important role in surveillance systems, which should be able to keep track of people after they have left the field of view of one camera and entered the field of view of any overlapping or non-overlapping camera. ReID accuracy can be significantly improved given a training set that demonstrates changes in appearances associated with the two nonoverlapping cameras involved. Here we test whether this advantage can be maintained when directly annotated training sets are not available for all camera-pairs at the site. Given the training sets capturing correspondences between cameras A and B and a different training set capturing correspondences between cameras B and C, the Transitive Re-IDentification algorithm (TRID) suggested here provides a classifier for (A,C) appearance pairs; see Fig. 1. TRID establishes a path between the non-directly trainable camera pair (A,C) by marginalization over the domain of possible appearances in camera B. Camera B plays the role of the ‘connecting element’ between cameras A and C. Our goal is to estimate the conditional probability P(YAC|xA,xC) where the notations xA and xC refer to a feature vector describing the appearance observed by cameras A and C and the binary variable YAC gets the value 1 if and only if the appearances given in A and in C are of the same identity. When the feature vector is known to correspond to a particular individual of identity i, we denote it by xi A. Thus, the pair {(xi A,x i C)} is a pair of feature vectors corresponding to the same person but to different cameras. Although a training set consisting of annotated pairs {(xi A,x i C)} is not available, we can exploit the annotated sets SAB = {(xi A,x i B)}, i = 1, ...,n and SBC = {(x j B,x j C)}, j = n+ 1, ...,n+m. The desired conditional probability is expressed by: