Research On Unsupervised Learning And Domain Adaptation In Person Re-Identification

Person re-identification aims to find pedestrian images with the same ID in a nonoverlapped camera system,which is crucial for security and new retail.With the rapid growth of deep neural networks,the accuracies of re-ID models have been greatly boosted and reached a remarkable level.However,these kinds of methods require a large amount of annotated data.At the same time,they are prone to perform well on settled scenes(source domain)but not for new scenes(target domain).These problems make the domain adaptation and unsupervised re-ID popular in recent papers.We propose ACT and PGPPM to solve the aforementioned problems.ACT makes good use of co-teaching-like training framework to mine reliable training samples from target domain for model refinement.To avoid local minimum during optimization,an asymmetric data-flow is proposed to make two sub-networks in ACT see different training samples and achieve better re-ID accuracy,PGPPM trains a generative model to overcome the difficulty of obtaining annotated samples.The unsupervised re-ID model is trained on generated virtual data and then adopted to mine positive training pairs with the proposed collaborative filtering-based mining algorithm for further boosting of the model.The main work and contributions of this paper are listed below:(1)To overcome the weakness of re-ID model in cross-domain accuracy,we proposed an unsupervised domain adaptation method called "ACT"(asymmetric co-teaching).ACT adopts mined training samples and co-teaching-like framework for further training of re-ID model to avoid the negative effects of noisy samples brought by clustering.To avoid restricting in local minimum during optimization,an asymmetric data flow is proposed and further improves re-ID accuracy.(2)To get rid of the demand of annotated data for training accurate re-ID models,we propose PGPPM,an unsupervised re-ID algorithm based on the generative model and mined positive training samples.PGPPM first leverages unlabeled images to train a generative model and then generate a labeled virtual dataset for the initialization of re-ID model.After that,a collaborative-filtering-based mining algorithm is proposed to mine training samples from the real dataset.Finally,mined real images and virtual images are sent to the re-ID model for higher accuracy.PGPPM makes real data getting involved in training,not only overcomes the weaknesses of generative models but further boosts re-ID's performance.ACT and PGPPM achieve state-of-the-art performance on several public datasets,which prove the effectiveness of our methods.