PPIRD: Patent-Product Image Retrieval Dataset

PPIRD is the dataset released with the NeurIPS 2025 paper:

Intermediate Domain Alignment and Morphology Analogy for Patent-Product Image Retrieval

PPIRD is designed for Patent-Product Image Retrieval (PPIR), where a model retrieves relevant patent images from a large patent gallery given a product image query. This setting is useful for studying patent infringement search, open-set image retrieval, cross-domain visual matching, and representation learning between product photos and patent line drawings.

• Project page: https://loslorien.github.io/idama-project/
• Paper: https://papers.nips.cc/paper_files/paper/2025/hash/154743e7e9688cf77db5ee75807bda82-Abstract-Conference.html
• Paper PDF: https://papers.nips.cc/paper_files/paper/2025/file/154743e7e9688cf77db5ee75807bda82-Paper-Conference.pdf
• Code: https://github.com/haifangong/IDAMA
• Dataset and models: https://huggingface.co/datasets/haifan-gong/PPIRD

Dataset Overview

Patent-Product Image Retrieval is challenging because product images and patent images differ substantially in visual style. Product images are usually RGB photographs, while patent images are often binary or grayscale line drawings. PPIRD provides a benchmark and supporting resources to study this cross-domain retrieval problem under an open-set setting.

The released resources include:

Component	Description
Test benchmark	Product–patent retrieval benchmark with annotated product–patent pairs
Retrieval gallery	Large-scale patent image pool used for retrieval
Unsupervised pre-training data	Product and patent images, including edge-domain variants for representation learning
Extracted features	Precomputed features used by the IDAMA pipeline
Model weights	Checkpoints and backbones used in the paper
Metadata	Test metadata and file-level indexing information

Dataset Scale

Split / Component	Scale	Description
Test pairs	439 product–patent pairs	Annotated product–patent pairs for retrieval evaluation
Retrieval pool	727,921 patent images	Large patent gallery used in the open-set retrieval setting
Unlabeled pre-training set	3,799,695 images	Large-scale product and patent images for representation learning
Edge-domain patent images	2,198,223 images	UAED-generated patent edge images
Edge-domain product images	1,601,458 images	UAED-generated product edge images
ImageNet-1K edge images	1,281,167 images	Additional edge-domain images derived from ImageNet-1K training images

Repository Structure

The Hugging Face repository is organized as follows:

PPIRD/
├── data/
│   ├── test/
│   │   ├── patent_A45.tar.gz
│   │   ├── patent_A47.tar.gz
│   │   ├── patent_A63.tar.gz
│   │   ├── patent_B65.tar.gz
│   │   ├── patent_H01.tar.gz
│   │   ├── product_all.tar.gz
│   │   └── test_metadata.tar.gz
│   │
│   └── unsupervised_train_edge/
│       ├── patent/
│       ├── product/
│       ├── imagenet1k-edge/
│       ├── manifest.jsonl
│       ├── summary.json
│       ├── explore_unsupervised_train.py
│       └── README.md
│
├── features/
│   └── ibot_features.tar
│
└── model/
    ├── checkpoints/
    ├── edge_feature/
    ├── eva/
    ├── eva02/
    ├── ibot/
    ├── mae/
    └── swin/

Data Description

Test Benchmark

The data/test/ directory contains the product queries, patent retrieval gallery, and test metadata.

The patent gallery is grouped by patent categories:

data/test/
├── patent_A45.tar.gz
├── patent_A47.tar.gz
├── patent_A63.tar.gz
├── patent_B65.tar.gz
├── patent_H01.tar.gz
├── product_all.tar.gz
└── test_metadata.tar.gz

The test benchmark is designed for open-set product-to-patent retrieval. Given a product image query, the model ranks patent images in the gallery. Correct matches are determined by the annotated product–patent pairs.

Unsupervised Edge-Domain Training Data

The data/unsupervised_train_edge/ directory contains edge-domain images used for unsupervised representation learning and intermediate-domain alignment.

data/unsupervised_train_edge/
├── patent/                 # Patent edge images, 29 tar shards
├── product/                # Product edge images, 21 tar shards
├── imagenet1k-edge/        # ImageNet-1K edge-domain images
├── manifest.jsonl          # Index for patent/product edge images
├── summary.json            # Dataset statistics
├── explore_unsupervised_train.py
└── README.md

The patent and product edge images are generated by the UAED edge detector. The manifest file provides file-level indexing information for streaming, lookup, and verification.

Each line in manifest.jsonl is a JSON object:

{
  "category": "patent",
  "shard_id": 0,
  "tar": "patent/shard_0000.tar",
  "member": "US10000225_1.jpg",
  "source": "/path/to/source/US10000225_1.jpg"
}

Field descriptions:

Field	Meaning
category	patent or product
shard_id	Shard index
tar	Relative path to the tar shard
member	Image filename inside the tar archive
source	Original source path before packaging

Download

You can download the full repository with huggingface_hub:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="haifan-gong/PPIRD",
    repo_type="dataset",
    local_dir="./PPIRD"
)

For partial download, use allow_patterns:

from huggingface_hub import snapshot_download

# Download only the test benchmark
snapshot_download(
    repo_id="haifan-gong/PPIRD",
    repo_type="dataset",
    local_dir="./PPIRD",
    allow_patterns=[
        "data/test/*"
    ]
)

# Download only the unsupervised edge-domain data
snapshot_download(
    repo_id="haifan-gong/PPIRD",
    repo_type="dataset",
    local_dir="./PPIRD",
    allow_patterns=[
        "data/unsupervised_train_edge/*"
    ]
)

# Download only model weights
snapshot_download(
    repo_id="haifan-gong/PPIRD",
    repo_type="dataset",
    local_dir="./PPIRD",
    allow_patterns=[
        "model/*"
    ]
)

Quick Start

Inspect the Unsupervised Edge-Domain Data

After downloading data/unsupervised_train_edge/, run:

cd PPIRD/data/unsupervised_train_edge

# Print dataset summary
python explore_unsupervised_train.py summary

# List tar shards
python explore_unsupervised_train.py list-shards

# List files inside one patent shard
python explore_unsupervised_train.py list-tar patent 0 --limit 10

# Randomly sample images for preview
python explore_unsupervised_train.py sample -o /tmp/ppird_preview -n 8 --category patent

# Verify a shard
python explore_unsupervised_train.py verify --category patent --shard 0

# Search the manifest by filename
python explore_unsupervised_train.py lookup --member US10000225_1

Extract One Shard

mkdir -p /tmp/patent_shard0
tar -xf patent/shard_0000.tar -C /tmp/patent_shard0

Extract a Subset for Debugging

python explore_unsupervised_train.py extract \
  --category patent \
  --shard 0 \
  --max-files 1000 \
  -o /tmp/ppird_subset

Streaming Usage

For large-scale training, we recommend reading tar shards directly instead of extracting all images to disk.

Example with Python tarfile:

import io
import tarfile
from PIL import Image

tar_path = "PPIRD/data/unsupervised_train_edge/patent/shard_0000.tar"

with tarfile.open(tar_path, "r") as tar:
    members = [m for m in tar.getmembers() if m.isfile() and m.name.endswith(".jpg")]
    member = members[0]
    f = tar.extractfile(member)
    image = Image.open(io.BytesIO(f.read())).convert("L")

print(member.name, image.size, image.mode)

Example with WebDataset-style streaming:

import webdataset as wds

dataset = (
    wds.WebDataset("PPIRD/data/unsupervised_train_edge/patent/shard_{0000..0028}.tar")
    .decode("pil")
    .to_tuple("jpg")
)

for image, in dataset:
    # image is a PIL image
    pass

Evaluation Protocol

PPIRD evaluates product-to-patent retrieval. Given a product image query, a method ranks all candidate patent images in the retrieval gallery. The annotated product–patent pairs are used as ground truth.

The official IDAMA codebase provides preprocessing, feature extraction, retrieval inference, and evaluation scripts:

https://github.com/haifangong/IDAMA

Please refer to the official repository for the exact evaluation pipeline and configuration used in the paper.

Baseline and Model Assets

This Hugging Face repository also includes model and feature files for reproducing the IDAMA experiments.

features/
└── ibot_features.tar

model/
├── checkpoints/
├── edge_feature/
├── eva/
├── eva02/
├── ibot/
├── mae/
└── swin/

These files are provided for research reproducibility. Please use the official GitHub repository for the corresponding loading and evaluation scripts.

Intended Uses

PPIRD is intended for research on:

• Patent-product image retrieval
• Open-set image retrieval
• Cross-domain visual matching
• Patent infringement search
• Product-to-line-drawing retrieval
• Edge-domain representation learning
• Self-supervised and unsupervised visual pre-training
• Intermediate-domain alignment

Out-of-Scope Uses

This dataset should not be used as the sole basis for legal decisions, infringement judgments, or commercial risk assessments. Patent infringement analysis requires expert legal interpretation and case-specific evidence beyond visual similarity.

Dataset Viewer Note

Because PPIRD is large and stored in tar/tar.gz archives, the Hugging Face dataset viewer may not display the full dataset. Please download the required files or use the provided scripts for local inspection.

Citation

If you find PPIRD or IDAMA useful, please cite:

@inproceedings{gong2025intermediate,
  title     = {Intermediate Domain Alignment and Morphology Analogy for Patent-Product Image Retrieval},
  author    = {Gong, Haifan and Zhang, Xuanye and Zhang, Ruifei and Su, Yun and Li, Zhuo and Du, Yuhao and Gao, Anningzhe and Wan, Xiang and Li, Haofeng},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {38},
  year      = {2025},
  url       = {https://papers.nips.cc/paper_files/paper/2025/hash/154743e7e9688cf77db5ee75807bda82-Abstract-Conference.html}
}

Contact

For questions, please open an issue in the official GitHub repository:

https://github.com/haifangong/IDAMA