ml-resources/MaritimeOCRBench

Mirror of HiDolphin/MaritimeOCRBench from ModelScope. All credit to the original authors. Not an official release from them. License unchanged (apache-2.0).

Maritime-OCR-Bench

简体中文

Overview

Maritime-OCR-Bench is an OCR and document understanding benchmark built around maritime-oriented tasks. In the maritime portion of the dataset, the samples mainly include container identification markings, bills of lading and shipping forms, port and cargo documents, and ship or engine manuals. Beyond these materials, the dataset also includes a broader mix of real-world documents, scanned pages, tables, receipts, invoices, reports, and scene-style text images.

Maritime-OCR-Bench is organized as an instruction-following multimodal conversation dataset. Each sample includes image paths, a user instruction, a reference answer, and a task type. The task design covers both free-form text understanding and strictly formatted outputs: VQA and Parsing emphasize semantic understanding, reading order, and long-text reconstruction, while IE and Spotting place more emphasis on JSON validity, field coverage, layout coordinates, and output-structure stability.

Task Types

This dataset includes 5 task types and 1888 samples.

Task	task_type in dataset	Samples	Description
IE	IE	471	Strict information extraction with JSON-formatted answers.
VQA	vqa	471	Visual question answering over cross-domain documents, tables, forms, receipts, and scene-style images.
Parsing	parsing	472	OCR parsing / full-text recognition, including long-form structured text.
JSON Spotting v1	json1	237	Text spotting with image-grouped JSON output for systems that consume results by page or image.
JSON Spotting v2	json2	237	Text spotting with unified images-based JSON output for systems that require a general downstream schema.

Task distribution in this dataset:

Data Format

The annotation file of this dataset uses JSONL format. Each sample is stored as one JSON object per line, and the top-level fields are:

• images: relative image path list.
• messages: multi-turn conversation containing the user instruction and the reference answer.
• task_type: task identifier.

Example:

{
  "messages": [
    {
      "role": "user",
      "content": "<image>Please extract all key information and return it in JSON format."
    },
    {
      "role": "assistant",
      "content": "{\n  \"field_a\": \"value\"\n}"
    }
  ],
  "images": [
    "images/example_image.jpg"
  ],
  "task_type": "IE"
}

File Structure

MaritimeOCRBench/
├── README.md
├── README.zh-CN.md
├── assets/
├── maritime_ocr_bench.jsonl
└── images/

The repository is organized into the following parts:

• README.md: English overview, task definitions, evaluation protocol, and reference results.
• README.zh-CN.md: Chinese overview, task definitions, evaluation protocol, and reference results.
• assets/: figures used in the README and supplementary documentation.
• maritime_ocr_bench.jsonl: the main annotation file of this dataset.
• images/: image assets referenced by the dataset annotation file.

Evaluation

Evaluation uses task-specific metrics:

• VQA / Parsing: a composite text score is used. Before scoring, the text is normalized and compared from four views: normalized text, compact text, loose text, and table-aware text. The final score is a weighted sum where edit-distance similarity, character-level F1, LCS F1, loose edit-distance similarity, and table-aware edit-distance similarity have weights of 0.25, 0.20, 0.15, 0.20, and 0.20 respectively. text_score = 0.25 * edit_sim_v1 + 0.20 * char_f1_v2 + 0.15 * lcs_f1_v1 + 0.20 * loose_edit_sim + 0.20 * table_aware_sim.
• IE: a combined score of text coverage and strict JSON validity is used. Text coverage measures how much of the reference-answer text is covered by the model output, without extra penalty for additional completion text. At the same time, the completion must be strict JSON text, with no extra explanation or Markdown code fences. ie_score = 0.5 * text_coverage_score + 0.5 * json_strict_score.
• Spotting Json1 / Json2: layout localization score and composite text score are computed separately and then aggregated with weights. overall_score = 0.7 * diou_score + 0.3 * text_score.

evalscope Integration

You can evaluate maritime_ocr_bench with the official modelscope/evalscope repository.

Example setup:

git clone https://github.com/modelscope/evalscope.git
cd evalscope

Then run the benchmark with:

from evalscope import TaskConfig, run_task

task_cfg = TaskConfig(
  model='qwen-vl-max',
  api_url='https://dashscope.aliyuncs.com/compatible-mode/v1',
  api_key='',
  datasets=['maritime_ocr_bench'],
  eval_batch_size=20,
  generation_config={
    'retries': 1,
    'retry_interval': 1,
    'timeout': 300,
  },
  ignore_errors=True,
  limit=None,
  debug=False,
)

run_task(task_cfg=task_cfg)

Reference Results

Full-set Results (1888 samples)

Model	Samples	Overall	IE	VQA	Parsing	Spotting_Json1 (2D array output)	Spotting_Json2 (1D array output)
Hi-Dolphin-OCR	1888	0.7941	0.7795	0.7817	0.8260	0.7700	0.8161