Рақамли технологияларнинг назарий ва амалий масалалари 9-tom 1-san (2026) · 49-56-betler
Evaluation of person re-identification methods based on four architectures of the SBS family
Хамдамов, Р.Х., Рахманов, Х.Э., Раматов, И.И.
Annotaciya
This paper considers the problem of person re-identification (Re-ID), which plays a key role in intelligent video surveillance and behavior analysis systems. The main focus is on the comparative analysis of methods implemented based on the Strong Baseline System (SBS), the basic neural network architecture for solving the Re-ID problem. The study analyzes different model variants: SBS(R50), which uses ResNet-50 as a base network; SBS(R50-IBN), which applies Instance-Batch Normalization to improve robustness to image style variations; SBS(S50), built on the basis of SE-ResNet-50 with a feature enhancement mechanism; and SBS(R101-IBN), which combines the deeper architecture of ResNet-101 with IBN to achieve maximum accuracy when working with large datasets. The analysis reveals the strengths and weaknesses of each SBS modification: from the high speed and stability of the base version (R50) to improved generalization ability (R50-IBN), enhanced discriminative features (S50), and high accuracy at high computational costs (R101-IBN). The results confirm that the choice of a specific model should depend on system requirements-speed, domain-shift tolerance, or a priority for maximum accuracy.
ReIDStrong Baseline System (SBS)SBS (R50)SBS (R50-IBN)SBS (S50)SBS (R101-IBN) YOLOv11нейросетевое извлечение признаковметрическое обучениевидеонаблюдениеre-identification of a person
Metadata derekkózi: jurnal OAI-PMH arxivi · Sindex tolıq mátindi saqlamaydı, derekkózge silteme beredi.