PHQPL: producing high quality pseudo labels for unsupervised person Re-IDentification

Unsupervised person Re-IDentification (Re-ID) algorithms typically employ complex large models as feature extractors, but they are frequently constrained in the physical deployment environments with finite computility resources. To address this concern, this paper proposes a novel producing high quality pseudo labels (PHQPL) framework for simple small network with the aim of achieving rapid inference speed while maintaining superior precision in the unsupervised person Re-ID task. Specifically, the novel PHQPL framework is primarily comprised of pseudo labels recycling (PLR) and soft pseudo labels creation (SPLC) module. The PLR module is designed to preserve the pseudo labels derived from the complex large (teacher) network as high quality supervisory information to guide the performance optimization of simple small (student) network through online pseudo label knowledge distillation operations. This module reduces the difficulty of eliminating the adverse implications of noisy pseudo labels upon model training by transforming unsupervised person Re-ID task with an agnostic proportion of noisy labels learning into a certain proportion of learning. The SPLC module adequately exploits the prior knowledge of teacher network and leverages its output values to produce soft pseudo labels that contain more hierarchical information about different person categories as high quality supervisory information for student network. Sufficient experiments are conducted on three person and one vehicle Re-ID benchmark datasets so as to authenticate the effectiveness of the methodology in this paper. The experimental findings demonstrate that the PHQPL framework allows the behavior of student network to approach or even surpass that of teacher network, exhibiting competitive outcomes over state-of-the-art algorithms.