README.md

---
license: other
task_categories:
- image-retrieval
- image-text-retrieval
- computer-vision
- natural-language-processing
language:
- en
tags:
- design-patent
- patent-retrieval
- patent-examination
- multimodal-learning
- vision-language-model
size_categories:
- 1K<n<10K
---

# PatMV: Multi-view Design Patent Retrieval Dataset

## Dataset Description

PatMV is a real-world multi-view design patent retrieval dataset constructed for research on **U.S. design patent novelty examination**.

Unlike conventional design patent retrieval benchmarks that generate positive pairs by splitting views from the same patent, PatMV is constructed from **official USPTO examination records**. Each query patent is paired with a prior design patent cited by professional patent examiners during novelty rejection procedures under **35 U.S.C. §102**.

The dataset aims to provide a realistic evaluation environment for automated design patent retrieval systems, especially for studying:

- multi-view design patent representation learning;
- cross-patent visual similarity retrieval;
- vision-language models for patent understanding;
- AI-assisted design patent examination.

The dataset contains examiner-confirmed similar design pairs and additional distractor patents to simulate realistic large-scale retrieval scenarios.

---

## Dataset Construction

PatMV was constructed from publicly available USPTO design patent examination documents.

The construction process includes:

1. **Mining USPTO examination records**

   Non-final rejection documents from USPTO were collected, and cases related to novelty rejection under 35 U.S.C. §102 were identified.

   Examiner-cited prior design patents were extracted to form query-prior patent pairs.

2. **Patent image collection and processing**

   Official patent drawing images were collected from USPTO records.

   The collected images were processed by removing irrelevant document information and manually verified to ensure data quality.

3. **Distractor construction**

   Additional design patents were introduced as distractors to simulate realistic retrieval scenarios.

4. **Multi-modal annotation**

   To support vision-language research, the dataset provides both image-level and patent-level textual descriptions.

   - Image-level captions are generated for individual patent views.
   - Patent-level captions summarize multiple views of the same patent.

---

## Dataset Statistics

The current benchmark contains:

| Component | Number |
|-----------|-------:|
| Query patents | 166 |
| Prior design patents (ground truth) | 161 |
| Distractor patents | 1610 |

Each query patent has one corresponding prior design patent in the gallery.

---

## Dataset Format

After extracting the dataset:

```

PatMV/
├── images/
│   ├── {patent_id}*view*{view_idx}.jpg
│   └── ...
│
├── query_metadata.json
│
└── gallery_metadata.json

```

### Images

The `images/` directory contains anonymized patent drawing images.

The naming format is:

```

{patent_id}*view*{view_idx}.jpg

````

where:

- `patent_id`: the identifier of the design patent;
- `view_idx`: the index of the patent view.

Each patent may contain multiple views, such as front, rear, side, top, bottom, or perspective views.

---

## Metadata Files

### query_metadata.json

This file contains metadata for query patents.

Each entry includes information such as:

- patent identifier;
- associated image views;
- image-level captions;
- patent-level caption.

Example:

```json
{
    "patent_id": "xxx",
    "images": [
        "xxx_view_0.jpg",
        "xxx_view_1.jpg"
    ],
    "image_captions": [
        "description of view 0",
        "description of view 1"
    ],
    "patent_caption": "overall description of the design"
}
````

---

### gallery_metadata.json

This file contains metadata for gallery patents, including:

* ground-truth prior designs;
* distractor patents;
* corresponding images;
* generated textual descriptions.

The gallery is designed for retrieval evaluation, where a query patent should retrieve its corresponding prior design from a collection containing distractors.

---

## Supported Research Tasks

PatMV can be used for the following research tasks:

### 1. Multi-view Design Patent Retrieval

Given a query patent containing multiple views, retrieve the corresponding prior design patent from a gallery.

### 2. Vision-Language Patent Understanding

Study the alignment between patent images and textual descriptions.

### 3. Design Similarity Analysis

Investigate computational methods for measuring visual similarity between different design patents.

### 4. AI-assisted Patent Examination

Develop intelligent systems to assist novelty analysis in U.S. design patent examination.

---

## Intended Use

PatMV is released for academic and research purposes.

Potential applications include:

* design patent retrieval;
* multimodal patent representation learning;
* patent image-text understanding;
* automated novelty search;
* AI-assisted patent examination systems.

---

## Limitations

PatMV focuses on U.S. design patent novelty examination cases available from public USPTO records.

The dataset is intended for research on computational methods for patent analysis and should not be interpreted as a legal determination of infringement.

Researchers should consider applicable intellectual property regulations and USPTO data usage policies when using this dataset.

---

## Citation

This dataset is introduced in the following work:

> Wang Yi-wei et al.
> "面向外观专利审查的专利级多视图检索方法."
> Accepted by CCIR 2026.

---

## License

This dataset is released for academic research purposes only.

Users are responsible for ensuring compliance with applicable laws, regulations, and intellectual property requirements.


---