An Exploration of Deep Learning Based Methods for Person Re-Identification

Abstract

Person Re-Identification (Re-ID) is an important assignment in the digital world where the access to the surveillance systems in a constrained or in-the-wild environment is possible. Person Re-ID in a video-based scenario is considered as state-of-the-art (SOTA) way of going forward to get fruitful results from this application in a non-overlapping surveillance feed scenario. In this thesis, four SOTA works are identified in terms of what architecture the respective authors have worked with and gained the maximum accuracy when compared with the accuracy gained by the same type of architectures. The provided hyper-parameters are changed to see what positive/negative effect different sets of hyper-parameters have on the SOTA architectures. Additionally, in this thesis an already used loss function in a different field is identified that can have positive/fruitful effects in Person Re-identification field. Following contributions are made in this work. (1) Four SOTA along-with a strong backbone network with pre-processed input ways have been identified. (2) Different sets of hyper-parameters are reported after extensive experimentation and generated better results than the original hyperparameters reported with the architecture. (3) A loss function is identified that is added in the SOTA, and the baseline architectures and the modified networks are trained again. The results are improved on MARS [9] dataset on which the SOTA are trained and reported. (4) The SOTA architectures are evaluated on an unseen dataset (P-DESTRE) with two different types of experiments performed (i) On Overall Dataset and (ii) On Cleaned Dataset and the accuracy of all SOTA and modified SOTA are reported in terms of mAP. The best performing model on MARS [9] dataset was modified MGH (added loss function) that reported a maximum of 83.61% mAP. On P-DESTRE, the best performing architecture on the cleaned dataset was the original MGH [5] that reported a maximum of 95.69% mAP. The trainings and evaluation are explained in their respective sections.