Person Re-Identification Via Deep Learning

In recent years,person re-identification has become a popular direction in the field of computer vision,and more and more researchers have been involved in the theory and applications of deep learning in person re-identification.However,complex environments such as illumination changes,background transformations,low-resolution images,occlusion,and similar dresses of different persons make the research on person re-identification still very challenging.Based on deep learning,this thesis dedicated to the deveoplement of new algorithms for person re-identification.The main contributions can be summarized as follows.(1)Learning diverse features is key to the success of person re-identification.Various part-based methods have been extensively proposed for learning local representations,which,however,are still inferior to the best-performing methods for person identification.We propose to construct a strong lightweight network architecture,termed PLR-OSNet,based on the idea of Part-Level feature Resolution over the Omni-Scale Network(OSNet)for achieving feature diversity.The proposed PLR-OSNet has two branches,one branch for global feature representation and the other branch for local feature representation.The local branch employs a uniform partition strategy for part-level feature resolution but produces only a single identity-prediction loss,which is in sharp contrast to the existing part-based methods.Empirical evidence demonstrates that the proposed PLR-OSNet achieves state-of-the-art performance on popular person Re-ID datasets,including Market1501,Duke MTMC-re ID and CUHK03,despite its small model size.(2)A big challenge of person re-identification(Re-ID)using a multi-branch network architecture is to learn diverse and discriminative features from the ID-labeled dataset.We seek to construct a multi-branch architecture with ensured feature diversity by employing a novel Slow-Drop Block(SDB)data augmentation method.The proposed SDB is to pick up a random block pattern and drop such a block pattern for several batches of input images during training.Since it may drop a large portion of any input image,this makes the training hard to converge.Hence,we propose a novel double-batch-split co-training approach for remedying this problem.In particular,we show that feature diversity can be well achieved with the use of multiple SDB branches by setting individual dropping ratio for each branch.Empirical evidence demonstrates that the proposed method achieves state-of-the-art performance on popular person Re-ID datasets,including Market1501,Duke MTMC-re ID and CUHK03.(3)Various tricks on training are systematically investigated for person re-identification.We focus on two main training configurations in the field of person re-identification,and three diverse network architectures and their variants are considered.We compared the influence of different settings in the same training mechanism and different training mechanisms on the performance of person re-identification model,and finally provide some useful suggestions on the training configuration for the model training of person re-identification.