2D Rotary Position Embedding for Scene Text Recognition with Transformers

Scene Text Recognition (STR) models often struggle with irregular text—such as curved, rotated, or perspective-distorted characters—because they rely on 1D positional encodings that fail to capture the 2D spatial relationships inherent in text images. This paper investigates whether adapting 2D Rotary Position Embeddings (RoPE) can improve the spatial reasoning of Transformer-based STR models without requiring architectural redesigns or significant parameter overhead.

The authors propose 2D-RoPE-STR, which modifies standard axial 2D-RoPE in two key ways to suit the STR domain. First, it uses an anisotropic dimension allocation, assigning more capacity to the row axis to reflect the typical aspect ratio of text. Second, it extends the rotary coupling into the decoder's cross-attention, allowing the autoregressive decoder to attend to encoder tokens based on their 2D spatial layout. This approach is essentially parameter-free, as it replaces fixed positional embeddings with rotation matrices, requiring only two learnable scalar parameters for per-axis scaling.

2D-RoPE-STR achieves state-of-the-art performance across six standard STR benchmarks, with an average accuracy of 90.4%. The model shows particularly strong gains on irregular text datasets like CUTE80 (curved) and SVTP (perspective-distorted). A diagnostic protocol, including controlled ablations and attention visualization, confirms that the 2D rotary formulation effectively enables the model to resolve character reading order in complex layouts where 1D baselines fail. The authors demonstrate that these improvements are achieved with minimal parameter cost compared to traditional learnable positional embeddings.

This work provides a robust, plug-and-play solution for improving spatial awareness in Transformer-based vision models. By demonstrating that relative 2D position can be encoded through simple rotations rather than learned tables, the paper offers a more efficient and effective way to handle the geometric diversity of real-world scene text, which is critical for applications like document digitization and autonomous navigation.