Fine-grained unsupervised tracklet matching

Abstract

Re-identification offers a useful tool for non-invasive biometric validation, surveillance, and human-robot interaction in a broad range of applications from crowd traffic management to personalised healthcare. Given the leaps and bounds that Artificial Intelligence has moved forward with applications almost everywhere, the tasks of Person Re-Identification and Vehicle Re-Identification still remain relatively unapplied. The performance of Deep Learning (DL) in the domain of Videos, due to it probably being the most challenging of fields to work with in the field of Computer Vision, still remains relatively low. However with the introduction of new DL based techniques in the field of Unsupervised Representation Learning, along with more video data than ever being created on a daily basis, work on the field is more in demand than ever before. In this project we aim to create a Dataset Pipeline that creates efficient task based datasets in an unsupervised manner, and create a mini dataset for the task of Person Re-Identification (Person Re-ID). We also aim at learning feature representations from these datasets using Unsupervised Representation Learning techniques such as Contrastive Learning, and then Transfer the feature learnings to the task of Person Re-ID.