A Study Of Continuous Person Re-Identification For Non-forgetting And High Generalization

Person re-identification(Re ID)aims to match target pedestrians across cameras and plays an important role in areas such as smart security,lost and found and unmanned supermarkets.Current mainstream Re ID works assume that all training data is available at the same time,using a one-off model training process.In practice,however,new data is generated daily in surveillance networks,meaning that training data is often available in batches.Following the setting of batch availability of training data,the efficient use of piecemeal new data for incremental training is called continuous learning person re-identification.Very few continuous learning person re-identification works currently exist and they are still in the preliminary exploratory stage.On the one hand,they are tested in a domain-independent manner,without considering a unified evaluation of the Re ID model’s non-forgetting and generalization capabilities.On the other hand,the data used in their training are from short-term datasets with small style differences,which do not take into account the complexity of real-life scenarios with large style differences before and after the data stream.These restricted settings,while reducing the difficulty of the study,do not meet the complex and diverse needs of realistic Re ID scenarios.Therefore,this paper proposes a new and more difficult test and training setup,namely person re-identification oriented towards non-forgetting and high generalization.The main contents of this paper are as follows :(1)To address the restricted test setup in the current work on continuous person re-identification,a setup for unified testing on a joint dataset(containing both seen and unseen data)is proposed.Under domain joint testing,the model additionally suffers from inter-domain negative sample interference,i.e.the inter-domain negative sample conflict problem.We design a continuous learning approach based on a mutual learning strategy for this purpose,which introduces a unified classifier setup while being able to balance the learning divergence of the Re ID model between the non-forgetting task and the generalization task,thus achieving a unified identity discrimination capability on the joint dataset.(2)To address the constrained training setup in the current work on continuous person re-identification,a setup that adds recently released long-term datasets to participate in training is proposed.In the mixed short-and long-term training settings,the model learns conflicting discriminatory knowledge during continuous learning,i.e.the problem of inter-domain knowledge conflict.We design a continuous learning approach based on a generalized knowledge learning strategy for this purpose,which aims to help the model learn generalized discriminative knowledge that is common before and after,so that it can effectively reduce the adverse effects of knowledge conflicts that exist before and after the data stream.In this paper,two different experimental setups are proposed for continu ous depth and generalized breadth,respectively,and the two methods designed above are tested correspondingly.The experimental results show that the continuous learning methods designed in this paper both exhibit less forgetting and higher generalization compared to the leading-edge works,and are almost optimal in all metrics.