Content-based image retrieval (CBIR) uses a representation of visual content to identify relevant images, and it is one of the fundamental research challenges extensively studied in the multimedia community for decades. There has been the importance of powerful features of convolutional neural networks utilized by CBIRs, but the pipeline of deep learning based unsupervised image retrieval is complicated. Despite all these efforts, there has not been a software library that covers these deep learning methods in a unified manner.
This paper presents ‘PyRetri‘ to fill this gap. ‘PyRetri’ is an open-source deep learning based unsupervised image retrieval toolbox built on PyTorch designed for software engineers and researchers. PyRetri is a flexible deep learning based unsupervised image retrieval toolbox designed with simplicity and flexibility in mind.
- Modular Design: There is a breakdown of the deep learning based unsupervised image retrieval into several stages, and users can easily construct an image retrieval pipeline by selecting and combining different modules.
- Flexible Loading: The framework is able to adjust to load several types of model parameters.
- Support of Multiple Methods: The framework directly supports several popular methods designed for deep learning based unsupervised image retrieval, which is also suitable for the re-identification of a person.
- Configuration Search Tool: The configuration search tool helps users to find the optimal retrieval configuration with various hyper-parameters.