Arbitrary-Shaped Text Detection via Ordered Construction of Convex Hull Vertices

Scene text detection enables the automation of numerous household functions, such as reading and interpreting labels and instructions, thereby facilitating smarter home management. Bottom-up text detection approaches, based on the composition of connected components, offer a direct method for text detection modeling. However, many of these methods involve complex post-processing steps, requiring intricate stitching processes to compute the visiting order of text components and avoid contour intersections. In this paper, we abandon the traditional rectangular text component design and instead utilize the point sets on the top and bottom edges of the text as the basic text components. Building on this foundation, we propose a novel bottom-up text modeling approach that treats the complex assembly of text components as a path-planning problem. By integrating the point sets on the top and bottom edges within a convex hull and designing a simple yet effective decoder, we accurately restore the visiting order of text contour. Consequently, the convex hull itself serves as the contour of the text, preventing the occurrence of contour intersections. This approach has achieved results that surpass the state-of-the-art on multiple text detection datasets.