The task of re-identification (ReID) is defined as the recognition of the same individual at different times and locations. State-of-the-art techniques share two very strong assumptions: the total number of people in the scene is known a priori, and there exists a total overlap of identity between a camera pair, that is, every person appears in both camera views. This is unrealistic for real-world re-identification scenarios, when there is no prior information about the same people reappearing in the scene at different views. We refer to this unconstrained setting as the 'open world' ReID problem. The open-world problem is more challenging for two reasons: (i) the total number of unique people within each camera and the scene as a whole (cross-cameras) are both unknown, and (ii) each subject may appear in some unknown subset of the cameras.In this paper we consider for the first time the most general openworld re-identification problem. To address this, we introduce a new Conditional Random Field (CRF) model, making three important contributions: (1) No label information is needed a priori, allowing the system to detect when a new person enters the camera network; (2) An 'open world' solver, that is, the model does not assume that a person will (re)appear in every camera; and (3) Producing a person count as a byproduct. Our approach provides generality that is lacking in existing state of the art closed world ReID solutions.The objective of the CRF is to assign the most likely correct assignment of multiple id labels simultaneously to all the nodes in the CRF. We assume as input a set ofA camera c i making the detection; the time of detection t i (we assume cameras are synchronized); the image position p i and velocity v i where the person was detected; and an appearance feature a i from the detection bounding box. The re-identification task is to correctly assign identity labelsTo address this task we propose a CRF G = {V, E}, where each node corresponds to a person detection (observation) V = {v i = x i }. Each edge corresponds to a similarity between nodes/persons E = {e i j = (v i , v j )}, and the label of each node corresponds to the identity of that person/detection. Our aim is to find the set of labels L that best fits all the observations X ,where U(l i |X ) and B(l i , l j |X ) denote unary and pairwise energy functions, respectively. Our algorithm proceeds in two steps, as explained in Algorithm 1. First, we solve the CRF allowing connections only between detections within the same camera. Second, we use that solution as an initial condition to build the connections between different cameras, creating the final CRF model. The structure and parameterisation of CRF at each stage is the same. We only increase the information included. To evaluate our contribution, we focus on the challenging SAIVTSoftbio database [1], that includes 150 people recorded using 8 different Step 2 Figure 1: CRF illustration. In the first step, only detections within the same camera are connected. In the second step, a restric...
