--- tags: - image-retrieval - benchmark - photobench - vision-language license: cc-by-nc-4.0 size_categories: - n<1K task_categories: - text-to-image language: - en - zh --- ## Quick Links

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

--- # PhotoBench PhotoBench is the first benchmark constructed from authentic, personal albums, designed to shift the paradigm from visual matching to personalized multi-source intent-driven photo retrieval. > **Leaderboard:** [PhotoBench Leaderboard](https://huggingface.co/spaces/SorrowTea/PhotoBench/) --- ## Dataset Description PhotoBench is an image retrieval benchmark with open-ended natural language queries. Unlike the protected version, PhotoBench gives you unrestricted access to the raw images, allowing you to use your own embedding models, caption generators, or agent-based retrieval workflows. This dataset contains: - **Test queries** for leaderboard submission (English + Chinese) - **Validation queries** with released ground truth for local self-evaluation - **Raw images** for all 3 albums (available upon request; not included in this repository due to size) | Album | Images | Test Queries | Validation Queries | |-------|--------|--------------|--------------------| | 1 | ~1,070 | 382 | 100 | | 2 | ~1,470 | 236 | 100 | | 3 | ~1,050 | 269 | 100 | | **Total** | | **887** | **300** | > **Note:** > - Use **`albumN_test.json`** if you want to submit to the [PhotoBench Leaderboard](https://huggingface.co/spaces/SorrowTea/PhotoBench/). Ground truth is hidden and evaluated on the server. > - Use **`albumN_validation.json`** if you want to evaluate your model locally. Ground truth is included in this file. --- ## Two Variants PhotoBench is released in two variants to support different research directions: | | PhotoBench (Full) | PhotoBench-Protected | |---|---|---| | **Images** | Raw original photos (~11 GB) | Not included | | **Features** | Use your own models (CLIP, SigLIP, etc.) | Pre-computed captions & embeddings provided | | **Metadata** | Extract your own (EXIF, timestamps, etc.) | Pre-computed metadata provided | | **Focus** | Unrestricted retrieval: embedding, caption, or agent | Agent planning only | | **Leaderboard** | [PhotoBench](https://huggingface.co/spaces/SorrowTea/PhotoBench/) | [PhotoBench-Protected](https://huggingface.co/spaces/SorrowTea/PhotoBench-Protected/) | | **Dataset** | [SorrowTea/PhotoBench](https://huggingface.co/datasets/SorrowTea/PhotoBench) | [SorrowTea/PhotoBench-Protected](https://huggingface.co/datasets/SorrowTea/PhotoBench-Protected) | - **PhotoBench (Full)** — For researchers who want to experiment with their own vision encoders, caption generators, or end-to-end agent pipelines. You get the raw images and complete freedom. - **PhotoBench-Protected** — For researchers focusing exclusively on **agent planning and reasoning**. No raw images are provided; you must work with pre-computed captions, embeddings, and metadata. This isolates the planning component from visual representation learning. --- ## Data Format ### Test Queries (`albumN_test.json`) For leaderboard submission. Each file is a JSON array of query objects: ```json [ { "query_cn": "摆满的书桌", "query_en": "cluttered desk" } ] ``` | Field | Type | Description | |------------|--------|------------------------------------------| | `query_cn` | string | Query text in Chinese | | `query_en` | string | Query text in English (primary language) | ### Validation Queries (`albumN_validation.json`) For local self-evaluation. Each file is a JSON array of query objects with released ground truth: ```json [ { "query_cn": "烧香的三姐妹", "query_en": "three sisters offering incense", "ground_truth": ["IMG_4906.JPG"] } ] ``` | Field | Type | Description | |----------------|----------|-----------------------------------------------| | `query_cn` | string | Query text in Chinese | | `query_en` | string | Query text in English (primary language) | | `ground_truth` | string[] | List of correct image filenames for this query | ```json [ { "query_cn": "摆满的书桌", "query_en": "cluttered desk" }, { "query_cn": "紫毛衣女孩", "query_en": "girl in purple sweater" } ] ``` | Field | Type | Description | |------------|--------|------------------------------------------| | `query_cn` | string | Query text in Chinese | | `query_en` | string | Query text in English (primary language) | ### Raw Images The raw images (`album1/`, `album2/`, `album3/`) contain the full-resolution original photos. **Total size:** ~11 GB **Format:** JPEG **Naming:** Original camera filenames (e.g., `IMG_1234.JPG`, `FullSizeRender.JPG`) > Raw images are not hosted in this repository due to size constraints. Please contact the authors or use the download instructions below. --- ## How to Use ### 1. Download Queries Both test and validation JSON files are available directly in this repository: ```bash # Via huggingface_hub CLI huggingface-cli download SorrowTea/PhotoBench-Full-HF --repo-type dataset --local-dir ./photobench ``` Or browse and download individual files from the **Files** tab above. - `test/albumN_test.json` — for leaderboard submission - `validation/albumN_validation.json` — for local self-evaluation ### 2. Download Raw Images Raw images are distributed separately. Contact the authors for access, or prepare the images according to the album structure: ``` raw_albums/ ├── album1/ │ ├── IMG_0001.JPG │ ├── IMG_0002.JPG │ └── ... ├── album2/ └── album3/ ``` ### 3. Build Your Retrieval System With the raw images and test queries, you can: - Extract image embeddings with any vision encoder (CLIP, SigLIP, etc.) - Generate captions with any VLM (GPT-4V, Qwen-VL, etc.) - Design multi-step agent workflows - Evaluate with your own metrics ### 4. Build the Submission File The dataset provides one `albumN_test.json` per album. Before submitting, you must **combine all albums into a single JSON array** and add the `album_id` field to each query object: **Step-by-step:** 1. Load `album1_test.json`, `album2_test.json`, and `album3_test.json`. 2. For each query object, add `"album_id": "1"` (or `"2"` / `"3"`). 3. Add a `"pred"` field containing the ordered list of predicted image filenames. 4. Merge all queries into one JSON array and save as `submission.json`. **Example transformation:** ```python import json submission = [] for album_id in ["1", "2", "3"]: with open(f"album{album_id}_test.json") as f: queries = json.load(f) for q in queries: submission.append({ "album_id": album_id, "query_en": q["query_en"], "pred": ["IMG_0001.JPG", "IMG_0002.JPG", ...] # your predictions }) with open("submission.json", "w") as f: json.dump(submission, f, indent=2) ``` **Submission format:** ```json [ { "album_id": "1", "query_en": "cluttered desk", "pred": ["IMG_1234.JPG", "IMG_5678.JPG", "IMG_9012.JPG"] } ] ``` Requirements: - `album_id`: `"1"`, `"2"`, or `"3"` (string). - `query_en`: Must match the test query **exactly** (case-sensitive). - `pred`: Ordered list of predicted image filenames. Order matters for NDCG. ### 5. Submit to Leaderboard Upload `submission.json` to the [PhotoBench Leaderboard](https://huggingface.co/spaces/SorrowTea/PhotoBench/). --- ## Evaluation The leaderboard computes the following metrics: | Metric | Description | |------------|----------------------------------------------------| | Recall@k | Proportion of ground-truth images in top-k | | NDCG@k | Normalized Discounted Cumulative Gain at rank k | Supported k values: **1, 5, 10, 20, 50, 100** Results are averaged per album, then averaged across albums for the final score. Only **full submissions** (all 3 albums, all queries) are eligible for public leaderboard ranking. --- ## Citation If you use PhotoBench in your research, please cite: ```bibtex @misc{photobench2026, title={PhotoBench}, year={2026}, eprint={2603.01493}, archivePrefix={arXiv}, primaryClass={cs.CV} } ``` --- ## License This dataset is released under the MIT License. --- ## Contact For questions or data access requests, please open an issue on this repository or contact the authors.