A Large-scale Dataset for Robust Complex Anime Scene Text Detection

TL;DR

This work tackles the gap in robust text detection for anime scenes by introducing AnimeText, a large-scale, multilingual dataset with 735k images and 4.2M text blocks. It employs a three-stage annotation pipeline, including a CLIP-H based hard negative classifier and multi-granularity / polygon annotations via SAM, to produce reliable, rich annotations. Cross-dataset benchmarks show significant domain gaps between natural and anime text, with AnimeText-trained detectors bridging this gap and achieving strong F1 scores on anime-specific tasks. The dataset and baselines establish a foundation for robust anime text localization and downstream multimodal language models, while future work aims to extend to dynamic scenes and full OCR benchmarks.

Abstract

Current text detection datasets primarily target natural or document scenes, where text typically appear in regular font and shapes, monotonous colors, and orderly layouts. The text usually arranged along straight or curved lines. However, these characteristics differ significantly from anime scenes, where text is often diverse in style, irregularly arranged, and easily confused with complex visual elements such as symbols and decorative patterns. Text in anime scene also includes a large number of handwritten and stylized fonts. Motivated by this gap, we introduce AnimeText, a large-scale dataset containing 735K images and 4.2M annotated text blocks. It features hierarchical annotations and hard negative samples tailored for anime scenarios. %Cross-dataset evaluations using state-of-the-art methods demonstrate that models trained on AnimeText achieve superior performance in anime text detection tasks compared to existing datasets. To evaluate the robustness of AnimeText in complex anime scenes, we conducted cross-dataset benchmarking using state-of-the-art text detection methods. Experimental results demonstrate that models trained on AnimeText outperform those trained on existing datasets in anime scene text detection tasks. AnimeText on HuggingFace: https://huggingface.co/datasets/deepghs/AnimeText

Figures (10)

• Figure 1: Multilingual text and confusing patterns examples in anime scene.
• Figure 2: Pipeline to building the AnimeText.
• Figure 3: Examples of missing samples and hard negative samples. Missing samples: Text instances that were not annotated during Stage 1 (false negatives); Hard negative samples: Background elements that were mistakenly identified as text during Stage 1 (false positives).
• Figure 4: Hierarchical annotations examples in AnimeText.
• Figure 5: Polygon annotations examples in AnimeText.