SVTRv2X: Scene Text Recognition Model Code and Evaluation Data

13.4 MB of code and likely evaluation data for the SVTRv2X model, a framework for Scene Text Recognition (STR). The dataset, authored by Jian Guo and last updated in June 2026, is shared under a CC-BY-4.0 license on figshare. It is associated with research addressing challenges in STR, such as sensitivity to spatial perturbations and handling diverse text styles.

Use Cases

• Benchmarking STR model robustness based on the described challenges with geometric distortions and complex backgrounds.
• Training or fine-tuning lightweight text recognition models based on the integration of modules like Jumble and Self-Distillation.
• Researching specialized model architectures for diverse text styles based on the described Mixture-of-Experts (MoE) Module.

Strengths

• The associated research paper describes a specific, state-of-the-art model architecture (SVTRv2X) with three complementary improvement modules.
• The dataset is shared under a permissive CC-BY-4.0 license, facilitating reuse and modification.
• The 13.4 MB size suggests a focused release of code and potentially compact evaluation corpora.

Limitations

• The dataset's exact content (e.g., specific images, annotations, or training data) is not detailed in the provided metadata.
• Row count and column-level documentation are absent, which limits suitability assessment for direct data analysis.
• The description focuses on the model's architecture and performance; the nature and scale of any included data files are unclear.

Provenance

Source

Jian Guo, via figshare.

Collection Method

Likely contains the source code and potentially evaluation datasets for the SVTRv2X model as described in the associated research.

Freshness

Last updated 2026-06-01 17:45:44