Open-Set Text Recognition: Concept, Dataset, Protocol, and Framework

This chapter gives a clear and detailed definition of the OSTR task. First, we describe the aim, goal, and scope, and formulate and define the OSTR task and the relation between OSTR and other tasks. Secondly, we narrow down to the specific protocols used to measure model performances on the task. Before reaching the specific protocol, we introduce the commonly used protocols and datasets to lay out the background. Finally, this chapter presents a general framework of OSTR as an implementation guideline to build models for the OSOCR task. Here, we first describe the overall design, including the modules, and how they interact with each other. Specifically, we discuss the training workflow of the framework, its testing workflow, and how it may cope with an open-world data stream in a production environment. We then formulate the key variables of individual modules, giving a brief introduction to the functionality. More implementations will be introduced in detail in later chapters. Open-Set Text Recognition: Concept, Dataset, and Protocol In addition, we explain the backward compatibility by examples, i.e., how the framework also fits zero-shot and close-set text recognition methods.