Yesianrohn/WATER-Data

1---2license: apache-2.03task_categories:4  - image-to-text5language:6  - en7tags:8  - scene-text-recognition9  - STR10  - OCR11  - artistic-text12  - wordart13  - synthetic-data14  - lmdb15size_categories:16  - 1M<n<10M17pretty_name: WATER-Data18configs: []19---20 21# WATER-Data: Datasets for WordArt-Oriented Scene Text Recognition22 23**WATER-Data** is the official dataset release for the paper24**"Advancing WordArt-Oriented Scene Text Recognition: Datasets and Methods" (ECCV 2026)**.25 26WordArt (artistic text) features highly customized fonts, textures, and layouts, making27**WA**ordArt-oriented scene **TE**xt **R**ecognition (**WATER**) substantially more challenging28than general Scene Text Recognition (STR). The primary bottleneck for WATER is the lack of29large-scale, stylistically diverse, and reliably annotated data. WATER-Data addresses this gap30by providing a large-scale synthetic suite, a carefully deduplicated real training set, and a31dedicated artistic-text benchmark.32 33- 📄 **Paper (arXiv):** https://arxiv.org/abs/2606.2448434- 💻 **Code:** https://github.com/YesianRohn/WATER35- 🧠 **Model code (OpenOCR-WATERec):** https://github.com/YesianRohn/OpenOCR-WATERec36- 🏋️ **Model weights:** https://huggingface.co/Yesianrohn/WATERec-Models37- 🖋️ **Artistic fonts:** https://huggingface.co/datasets/Yesianrohn/artistic-fonts38- 📝 **WATER-Z captions:** https://huggingface.co/datasets/Yesianrohn/WATER-Z_Captions39 40---41 42## Dataset Overview43 44WATER-Data contains three components: a synthetic training suite (**WATER-S**), a real training45set (**WATER-R**), and an artistic-text evaluation benchmark (**WordArt-Bench**).46 47| Component | Subset | Role | #Instances | Source |48|-----------|--------|------|-----------|--------|49| **WATER-S** | WATER-T | Synthetic train | ~1M | Tool-based rendering (SynthWordArt) |50| **WATER-S** | WATER-Z | Synthetic train | ~1M | Generative model (Qwen3-VL + Z-Image) |51| **WATER-R** | – | Real train | 3,225,130 | Union14M-L + WordArt-Train + WAS-R (deduplicated) |52| **WordArt-Bench** | – | Evaluation | 1,511 | WordArt test split |53 54All subsets are English WordArt in the current release.55 56---57 58## Directory Structure59 60Every split is stored as a standalone **LMDB** database (`data.mdb` + `lock.mdb`), the format61used by the [OpenOCR](https://github.com/Topdu/OpenOCR) framework.62 63```64WATER-Data/65├── WATER-R/                 # Real training set (~11.8 GB)66│   ├── data.mdb67│   └── lock.mdb68├── WATER-S/                 # Synthetic training suite69│   ├── WATER-T/             # Tool-rendered subset70│   │   ├── data.mdb71│   │   └── lock.mdb72│   └── WATER-Z/             # Model-generated subset73│       ├── data.mdb74│       └── lock.mdb75└── WordArt-Bench/           # Artistic-text benchmark (~325 MB)76    ├── data.mdb77    └── lock.mdb78```79 80---81 82## Subset Details83 84### WATER-S — Synthetic Suite (≈2M)85A 2M-scale synthetic artistic-text dataset, improving the scale of existing artistic text data86by hundreds of times. It consists of two complementary subsets:87 88- **WATER-T (Tool-based Rendering, ~1M).** Generated with **SynthWordArt**, an artistic-text89  rendering engine built on SynthText / SynthTIGER. It replaces standard fonts with a library of90  **11,250 artistic fonts** and adds rich layout patterns (curved lines, vertical text,91  multi-orientation layouts, perspective and stretching). It offers **precise control** over text92  content, font, and layout, with perfectly accurate labels.93- **WATER-Z (Model-based Generation, ~1M).** Generated by an automatic few-shot prompt-mining94  pipeline: **Qwen3-VL-8B** mines fine-grained captions (with an editable text placeholder) from95  real artistic text, expands them into **273,488 high-quality prompts**, and **Z-Image-Turbo**96  synthesizes images at 256×256. It offers **higher realism and diversity** in background texture,97  layout composition, and global visual style.98 99WATER-T and WATER-Z are complementary: WATER-T provides strong controllability and label accuracy,100while WATER-Z provides natural, design-like style diversity. Training on their combination covers101both the "strongly controlled" and "style-diverse" regimes.102 103### WATER-R — Real Training Set (3.2M)104A real-world training set re-constructed from three sources:105[Union14M-L](https://github.com/Mountchicken/Union14M),106[WordArt-Train](https://github.com/xdxie/WordArt), and107[WAS-R](https://github.com/xdxie/WordArt). **Strict hashing deduplication** is performed against all108evaluation sets to avoid label leakage. It contains **3,225,130** text instances.109 110### WordArt-Bench — Evaluation Benchmark111The artistic-text evaluation benchmark (test split of WordArt), with **1,511** images, used to112report recognition accuracy. In the paper, our WATERec baseline reaches **90.40%** accuracy on this113benchmark — the first result to exceed 90% — surpassing both general-purpose and OCR-specialized114vision-language models by a large margin.115 116---117 118## Usage119 120Each LMDB database stores image–label pairs in the OpenOCR convention. A minimal reading example:121 122```python123import lmdb124 125env = lmdb.open(126    "WATER-Data/WordArt-Bench",  # folder containing data.mdb / lock.mdb127    readonly=True, lock=False, readahead=False, meminit=False,128)129 130with env.begin(write=False) as txn:131    num_samples = int(txn.get(b"num-samples"))132    # keys follow the OpenOCR layout, e.g.:133    #   image-000000001 -> raw image bytes134    #   label-000000001 -> ground-truth text135    img_buf = txn.get(b"image-000000001")136    label = txn.get(b"label-000000001").decode("utf-8")137 138print(num_samples, label)139```140 141For training and evaluation, we recommend using the official framework142[OpenOCR-WATERec](https://github.com/YesianRohn/OpenOCR-WATERec), which consumes these LMDB143databases directly.144 145To download the dataset:146 147```bash148# Requires: pip install -U "huggingface_hub[cli]"149hf download Yesianrohn/WATER-Data --repo-type dataset --local-dir ./WATER-Data150```151 152---153 154## Intended Use155 156WATER-Data is intended for **research** on scene text recognition, especially artistic / WordArt157text. Typical uses include: training and benchmarking STR models, studying synthetic-data158strategies (tool-based vs. generative), and evaluating general / OCR-specialized VLMs on159challenging stylized text.160 161---162 163## License164 165Released under the **Apache 2.0** license. The dataset is built upon publicly available STR data166sources (Union14M-L, WordArt, WAS-R) and synthetic content; please also respect the original167licenses of these underlying datasets.168 169---170 171## Citation172 173If you use WATER-Data in your research, please cite our paper:174 175```bibtex176@inproceedings{water2026eccv,177  title     = {Advancing WordArt-Oriented Scene Text Recognition: Datasets and Methods},178  author    = {Ye, Xingsong and Du, Yongkun and Zhang, Jiaxin and Zhang, Haojie and Sun, Chong and Li, Chen and Lyu, Jing and Chen, Zhineng},179  booktitle = {European Conference on Computer Vision (ECCV)},180  year      = {2026}181}182```183