README.md

---
language:
- en
license: mit
task_categories:
- other
tags:
- image-retrieval
- multimodal-agents
- visual-history
---

# DISBench: DeepImageSearch Benchmark

[**Paper**](https://huggingface.co/papers/2602.10809) | [**GitHub**](https://github.com/RUC-NLPIR/DeepImageSearch) | [**Leaderboard**](https://huggingface.co/spaces/RUC-NLPIR/DISBench-Leaderboard)

DISBench is the first benchmark for context-aware image retrieval over visual histories. It contains 122 queries across 57 users and 109,467 photos, requiring multi-step reasoning over corpus-level context.

## Download

**Option A: Hugging Face (Recommended)**

```bash
huggingface-cli download RUC-NLPIR/DISBench --local-dir DISBench
```

**Option B: Manual Download**

```bash
python download_images.py --photo-ids-path photo_ids --images-path images
```

## Sample Usage

To use DISBench with the ImageSeeker agent framework, you can follow these steps from the official repository:

### 1. Installation
```bash
git clone https://github.com/RUC-NLPIR/DeepImageSearch.git
cd DeepImageSearch
pip install -r requirements.txt
```

### 2. Run the Agent
```bash
python ImageSeeker/run_agent.py \
    --dataset_path DISBench \
    --model_name "claude-opus-4-5-20251101" \
    --api_base_url "YOUR_API_BASE_URL" \
    --api_key "YOUR_API_KEY" \
    --embedding_model "qwen3-vl-embedding-8b" \
    --max_turns 30 \
    --output_dir results/claude-opus-4.5/
```

## File Structure

```
DISBench/
├── queries.jsonl              # 122 annotated queries
├── metadata/
│   └── {user_id}.jsonl        # Photo metadata per user
├── images/
│   └── {user_id}/
│       └── {photo_id}.jpg     # Photo files
├── photo_ids/
│   └── {user_id}.txt          # Photo IDs and hashes per user
└── download_images.py         # Image download script
```

## Data Format

### queries.jsonl

Each line is a JSON object representing one query:

```json
{
    "query_id": "1",
    "user_id": "10287726@N02",
    "query": "Find photos from the musical performance identified by the blue and white event logo on site, where only the lead singer appears on stage.",
    "answer": ["7759256930", "7759407170", "7759295108", "7759433016"],
    "event_type": "intra-event"
}
```

| Field | Type | Description |
|:------|:-----|:------------|
| `query_id` | string | Unique query identifier |
| `user_id` | string | User whose photo collection to search |
| `query` | string | Natural language query (text-only) |
| `answer` | list[string] | Ground-truth target photo IDs |
| `event_type` | string | `"intra-event"` or `"inter-event"` |

### metadata/{user_id}.jsonl

Each line is a JSON object representing one photo's metadata:

```json
{
    "photo_id": "4517621778",
    "metadata": {
        "taken_time": "2010-04-10 13:52:57",
        "longitude": -1.239802,
        "latitude": 51.754123,
        "accuracy": 16.0,
        "address": "Y, Cherwell Street, St Clement's, East Oxford, Oxford, Oxfordshire, England, OX4 1BQ, United Kingdom",
        "capturedevice": "Panasonic DMC-TZ5"
    }
}
```

| Field | Type | Description |
|:------|:-----|:------------|
| `photo_id` | string | Unique photo identifier |
| `metadata.taken_time` | string | Capture time in `YY-MM-DD HH:MM:SS` format |
| `metadata.longitude` | float | GPS longitude. **Missing if unavailable.** |
| `metadata.latitude` | float | GPS latitude. **Missing if unavailable.** |
| `metadata.accuracy` | float | GPS accuracy level. **Missing if unavailable.** |
| `metadata.address` | string | Reverse-geocoded address. **Missing if unavailable.** |
| `metadata.capturedevice` | string | Camera/device name. **Missing if unavailable.** |

> **Note:** Optional fields (`longitude`, `latitude`, `accuracy`, `address`, `capturedevice`) are omitted entirely when unavailable — they will not appear as keys in the JSON object.

### photo_ids/{user_id}.txt

Each line represents one photo ID and its hash on aws storage in the format `{photo_id}\t{hash}`:
```
1205732595	c45044fd7b5c9450b2a11adc6b42d
```

## Dataset Statistics

| Statistic | Value |
|:----------|:------|
| Total Queries | 122 |
| Intra-Event Queries | 57 (46.7%) |
| Inter-Event Queries | 65 (53.3%) |
| Total Users | 57 |
| Total Photos | 109,467 |
| Avg. Targets per Query | 3.84 |
| Avg. History Span | 3.4 years |
| Query Retention Rate | 6.1% (122 / 2,000 candidates) |
| Inter-Annotator IoU | 0.91 |

## Data Source

DISBench is constructed from [YFCC100M](https://multimediacommons.wordpress.com/yfcc100m-core-dataset/), which preserves a hierarchical structure of users → photosets → photos. All images are publicly shared under Creative Commons licenses. Photoset boundaries are used during construction but are **not** provided to models during evaluation.

## Citation

```bibtex
@misc{deng2026deepimagesearchbenchmarkingmultimodalagents,
  title={DeepImageSearch: Benchmarking Multimodal Agents for Context-Aware Image Retrieval in Visual Histories}, 
  author={Chenlong Deng and Mengjie Deng and Junjie Wu and Dun Zeng and Teng Wang and Qingsong Xie and Jiadeng Huang and Shengjie Ma and Changwang Zhang and Zhaoxiang Wang and Jun Wang and Yutao Zhu and Zhicheng Dou},
  year={2026},
  eprint={2602.10809},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2602.10809}
}
```

## License

The DISBench dataset follows the Creative Commons licensing terms of the underlying YFCC100M data. Please refer to individual image licenses for specific usage terms.