Robust Knowledge Adaptation for Federated Unsupervised Person ReID

Abstract

Person Re-identification (ReID) has been extensively studied in recent years due to the increasing demand in the public security sector. However, collecting and modelling with sensitive personal data raises privacy concerns.

Therefore, federated learning has been studied for Person ReID, which aims to share minimal sensitive data between different parties (clients).

However, the statistical heterogeneity between client domains under a federated setting remains as a challenge, which limits the process of knowledge aggregation across clients and often leads to inferior identification accuracy.

Additionally, existing federated learning-based person ReID methods generally rely on laborious and time-consuming data annotations, which has the scalability issues to deploy them for real-world Person ReID.

Therefore, this thesis aims to address the unsupervised person ReID problem under a federated learning scheme. Specifically, two methods are devised:

1.In Chapter 3, we introduce a federated unsupervised cluster-contrastive (FedUCC) learning method for Person ReID. FedUCC introduces a three-stage modelling strategy following a coarse-to-fine manner. In detail, generic knowledge, specialized knowledge and patch knowledge are discovered using a deep neural network. This enables mutual knowledge sharing among clients while retaining local domain-specific knowledge based on the categories of the network components and their parameter settings.

2.In Chapter 4, we propose a novel deep transformer-based architecture - context and camera invariant transformer, namely CCIT, in pursuit of homogeneous image representations by ignoring the irrelevant context and camera bias.