Abstract

Abstract: It is important to perform real time search while Large scale Personre-identification on a large gallery. The conventional methods use to focus ondiscriminative learning which is probabilistic approach of learning. In thisproposed project we attempt to use deep learning while integrating it withhashing which provides a framework to evaluate productivity, precision and reliabilityof Large scale Person re-identification. We use augmentation for creatingartificial training images through different ways of processing or combinationof multiple processing such as random rotation, shifts, shear and flips etc. Wepropose Part based deep Hashing (PDH) in which augmented images are the  input of deep learning architecture. All theaugmented images have different identity. We are using whole image and use itfor training deep hashing architecture. We use a ternion loss function whichcalculates the hamming distance of the pedestrian image. The hamming distanceof the images with same identity is smaller than the one with differentidentity. In this project we use standard large scale dataset specificallyMarket-1501 & Market-1501 +500K.

Keywords: Deep learning, hashing, part-based, large-scale personre-identification.