Title: SynthBendText3D: a framework for generating scene text data in arbitrary orientations using a 3D graphics engine
Authors: Zhao Guan; Weilong Zhang
Addresses: School of Cyberspace Security and Computer, Hebei University, Baoding – 071000, China ' School of Cyberspace Security and Computer, Hebei University, Baoding – 071000, China
Abstract: To address the domain distribution mismatch between synthetic scene text data and real-world scene text data in arbitrary orientations, we introduce SynthBendText3D - a framework based on a 3D graphics engine that synthesises scene text data in various orientations. The framework generates a large number of text instances in arbitrary directions and constructs a 3D scene to position these instances. By leveraging domain randomisation techniques, it randomises scene parameters such as object arrangement, materials, lighting, and camera angles, ensuring a high degree of diversity in the synthesised data. Moreover, the framework incorporates a polygon reconstruction algorithm to annotate each synthesised text instance with polygonal bounding boxes. Experimental results demonstrate the effectiveness of the data generated by our framework.
Keywords: scene text detection; synthetic data; domain randomisation; domain adaption.
DOI: 10.1504/IJICT.2025.144017
International Journal of Information and Communication Technology, 2025 Vol.26 No.1, pp.38 - 54
Received: 08 Oct 2024
Accepted: 24 Oct 2024
Published online: 20 Jan 2025 *