Anchoring Regularization For Vehicle Re-Identification

Vehicle re-identification (re-ID) aims to automatically findvehicle identity from a large number of vehicle images cap-tured from multiple cameras. Most existing vehicle re-IDapproaches rely on fully supervised learning methodologies,where large amounts of annotated training data are required.In practice, massive data annotation is an expensive task andmay be impossible with real time learning and identification,in which, semi or unsupervised learning is needed. In this pa-per, we focus our interest on semi-supervised vehicle re-ID,where each identity has a single labeled and multiple unla-beled samples in the training. We propose a framework whichgradually labels vehicle images taken from surveillance cam-eras. Our framework is based on a deep Convolutional NeuralNetwork (CNN), which is progressively learned using a fea-ture anchoring regularization process. The experiments con-ducted on various publicly available datasets demonstrate theefficiency of our framework in re-ID tasks. Our approachwith only 20% labeled data shows interesting performancecompared to the state-of-the-art supervised methods trainedon fully labeled data.