README.md

metadata

configs:
  - config_name: samples
    data_files:
      - split: test
        path:
          - samples/*/query.json
  - config_name: protected
    data_files:
      - split: test
        path:
          - protected/*/test.json
  - config_name: validation
    data_files:
      - split: test
        path:
          - protected/*/validation.json
task_categories:
  - text-to-image
paper_ids:
  - '2603.01493'

Quick Links

🏠 GitHub · 📄 arXiv · 🏅 Leaderboard · 📦 Dataset · 🏅 Protected Leaderboard · 📦 Protected Dataset · 🖼️ Raw Images: obox · Google Drive (pwd: Oppo2026)

---

PhotoBench-Protected Dataset Overview

PhotoBench-Protected is the limited-information release of PhotoBench, the first benchmark constructed from authentic, personal albums for personalized multi-source intent-driven photo retrieval. This release provides pre-computed captions, embeddings, and metadata for agent-based evaluation.

This folder provides two components for evaluation and research:

• Samples: a small, display-oriented subset with images for demonstration only.
• Protected: machine-readable features without raw images for local evaluation.

Directory Layout

dataset/
├── README.md            # English overview (this file)
├── README_CN.md         # Chinese overview
├── samples/             # Display subset (per album, 20 queries with display images)
│   ├── album1/
│   │   ├── query.json   # 20 queries with GroundTruth image filenames
│   │   └── images/      # display images referenced by GroundTruth
│   ├── album2/
│   └── album3/
│   ├── README.md
│   └── sample_queries.schema.json
└── protected/           # Features for local evaluation (no raw images)
    ├── album1/
    │   ├── test.json        # test queries for leaderboard evaluation
    │   ├── validation.json  # validation queries for local self-testing
    │   ├── query.json       # all queries (test + validation combined)
    │   └── images/
    │       ├── captions/    # text indices by model
    │       ├── embeddings/  # vector indices by model
    │       ├── metadata/
    │       │   └── metadata.json  # location/time metadata (privacy-preserving)
    │       └── faceid/
    │           ├── face_info_cn.json
    │           ├── face_info_en.json
    │           └── face_id_images/
    ├── album2/
    └── album3/

Test vs. Validation Split

To prevent data leakage and overfitting, the protected release strictly separates test and validation queries:

• test.json — Queries used for official leaderboard evaluation. Submit predictions for these queries to the Protected Leaderboard.
• validation.json — Queries for local self-testing only. These will not be scored on the leaderboard.
• query.json — Combined set of all queries (test + validation) for convenience.

Why this split? The full PhotoBench dataset provides validation.json with ground-truth annotations so you can tune your model locally. To ensure fair comparison, the protected leaderboard only evaluates on test.json queries. Do not train or tune on test queries.

Files and Field Definitions

protected/test.json

• Type: list of query objects for leaderboard evaluation.
• Key fields:
  • query_cn: text of the retrieval intent (Chinese).
  • query_en: text of the retrieval intent (English).
  • Location, Time, Person, Object, Concept, Genre: dimension annotations.
  • Source: source category label.

protected/validation.json

• Same structure as test.json.
• For local self-evaluation only; not evaluated on the leaderboard.

protected/query.json (album-level, all queries)

• Combined test + validation queries.
• Contains the same query fields as test.json.

protected/images/captions/

• Per-model text indices for caption-based retrieval.
• Each metadata.json includes:
  • model_name: the name of the caption model.
  • filenames: list of image filenames included in the index.

protected/images/embeddings/

• Per-model vector indices for image embedding retrieval.
• Each metadata.json includes:
  • model_name
  • filenames

protected/images/metadata/metadata.json

• Privacy-preserving location and time metadata for images.
• Fields:
  • latitude: decimal degrees rounded to approximately 10 meters.
  • longitude: decimal degrees rounded to approximately 10 meters.
  • time: ISO-format timestamp string.

protected/images/faceid/

• Anonymized person information to support social/identity reasoning without exposing raw identities.
• Files:
  • face_info_cn.json / face_info_en.json include:
    • face_id_to_nicknames: mapping from face IDs to anonymized names/nicknames.
    • image_to_face_ids: mapping from image filename to a list of face IDs present.
  • face_id_images/: representative images for face IDs that appear.

Intended Use and Evaluation

• Samples: for display and qualitative understanding; not for model training.
• Protected: for local algorithms to run against machine-readable features.
• Evaluation: submit results (e.g., submission.json) for scoring against hidden GroundTruth.

Submission Format

When submitting to the leaderboard, combine queries from all albums into a single JSON array and add the album_id field to each query object:

[
  {
    "album_id": "1",
    "query_en": "cluttered desk",
    "pred": ["IMG_1234.jpg", "IMG_5678.jpg", ...]
  },
  {
    "album_id": "2",
    "query_en": "girl in purple sweater",
    "pred": ["IMG_9012.jpg", ...]
  }
]

Important:

• album_id must be "1", "2", or "3".
• query_en must match exactly (case-sensitive).
• pred is an ordered list of predicted image filenames. Order matters for NDCG.
• Only full submissions (all 3 albums, all test queries) are eligible for public leaderboard ranking.