MSIR-Bench Review Dataset

This repository contains an anonymized review snapshot of MSIR-Bench, a benchmark for identity-preserving style image retrieval.

Dataset Description

Each source identity is represented by an anonymous five-digit ID. Images are organized by split and identity folder. File names follow either <id>_<Style>.png, <id>_original.png, or legacy original.jpg for original reference images.

The dataset is intended for evaluating whether a retrieval model can match images by source identity across visual styles.

Splits

Styles

Seen styles include:

Impressionism, LineArt, InkWash, Gongbi, Ukiyoe, PencilSketch, RoughSketch, Crayon, PopArt, Claymation, Embroidery, Pixar, DigitalPainting, LowPoly, PixelArt, Glitch, Anime, Comic, Caricature.

Held-out or generated evaluation styles include project-defined unseen style candidates such as VintagePhoto, Vaporwave, StreetArt, Bauhaus, ClassicOil, and ChineseCartoon.

Privacy and Anonymization

• Identity folders use anonymous numeric IDs only.
• JPEG images were re-encoded before upload to remove EXIF fields, camera metadata, thumbnails, local paths, and author/tool-specific metadata when present. PNG files are distributed under anonymous numeric paths; PNG does not use EXIF in the same way as camera JPEGs.
• The released package avoids machine-specific absolute paths.
• No personal names or demographic labels are included in the metadata.

Important limitation: anonymization removes file and embedded metadata, but it does not alter visible image content. Users should treat the images as research data and avoid attempts to identify any depicted person or source.

Files

• data/: anonymized, privacy-processed images.
• metadata.csv: one row per image with relative path, split, identity ID, style, and file extension.
• dataset_statistics.json: computed image and identity statistics.
• supplementary/: path-anonymized code, split IDs, configs, and project metadata for review.

Usage

from datasets import load_dataset

dataset = load_dataset("imagefolder", data_dir="data")

For project-specific evaluation protocols, see supplementary/README.md and scripts under supplementary/code/.

Intended Use

This dataset is for anonymous academic review, research reproduction, and benchmarking of identity-preserving style retrieval methods.

Out-of-Scope Use

Do not use this dataset for face recognition deployment, biometric identification, surveillance, re-identification, or attempts to infer sensitive personal attributes.

Citation

Citation information is omitted during anonymous review and should be added after de-anonymization.