README.md

---
language:
- en
pretty_name: SnapBench
task_categories:
- visual-document-retrieval
tags:
- multimodal
- image-text-retrieval
- retrieval
- robustness
- benchmark
configs:
- config_name: queries
  data_dir: queries
  default: true
- config_name: gallery
  data_dir: gallery
---

Source repository: <https://github.com/zrchen03/SnapBench>

Paper: [SnapBench: Benchmarking Snap-and-Ask Multimodal Retrieval for Mobile Interactions](https://arxiv.org/abs/2608.29607)

# SnapBench: Benchmarking Snap-and-Ask Multimodal Retrieval for Mobile Interactions

SnapBench is a benchmark for **snap-and-ask** mobile interactions: a user captures a photo and asks a short English question. Each query pairs an image with text; the gallery contains image–caption pairs to retrieve from. The benchmark includes a clean split plus text and image perturbations to simulate real-world query degradation.

| | |
|---|---|
| Queries | 1,145 (image + text) |
| Gallery | 9,085 items (image + caption) |
| Conditions | 54 (1 clean + 8 text + 45 image perturbations) |

This repository ships the clean benchmark in the standard Hugging Face `ImageFolder` format. It includes the clean query and gallery images, retrieval annotations, and perturbation metadata. It does not include evaluation code.

---

## What Is Included

| Component | Location | Status |
|---|---|---|
| Query metadata | `queries/test/metadata.jsonl` | included |
| Query images | `queries/test/images/` (1,145) | included |
| Gallery metadata | `gallery/test/metadata.jsonl` (9,085 items) | included |
| Gallery images | `gallery/test/images/` (9,059 unique images) | included |
| Text perturbations | `queries/test/metadata.jsonl` → `text_perturbations` | included |
| Image perturbation specifications | `queries/test/metadata.jsonl` → `image_perturbations` | included |
| Image perturbation files | 15 types × 3 severity levels × 1,145 queries | **generate locally from the source repository** |

---

## Setup

Install Hugging Face Datasets:

```bash
pip install datasets
```

Load the two dataset configurations from the Hub:

```python
from datasets import load_dataset

queries = load_dataset("yefd/SnapBench", "queries", split="test")
gallery = load_dataset("yefd/SnapBench", "gallery", split="test")
```

To load a local copy of this repository:

```python
from datasets import load_dataset

queries = load_dataset(".", "queries", split="test")
gallery = load_dataset(".", "gallery", split="test")
```

The `file_name` field in each `metadata.jsonl` file is automatically exposed as an `image` column by the Hugging Face `ImageFolder` builder.

---

## Build the Full Benchmark

This Hugging Face version already includes the clean benchmark (queries, gallery, and text perturbations). The image perturbation specifications are stored in the `image_perturbations` field of every query. Use the scripts in the source repository to generate all image-perturbed query images locally.

### Step 1. Generate image perturbations

Clone and set up the source repository:

```bash
git lfs install
git clone https://github.com/zrchen03/SnapBench.git SnapBench_raw
cd SnapBench_raw
git lfs pull
pip install -r requirements.txt
export BENCH_IMAGES_DIR=$(pwd)/bench_images
```

Generate 15 perturbation types × 3 severity levels (sev1 / sev2 / sev3) × 1,145 queries = **51,675 images**:

```bash
python benchmark/gen_image_perturbations.py
```

Output:

```
bench_images/perturbed/{type}/sev{1,2,3}/{query_id}.jpg
```

Preview the workload without writing files:

```bash
python benchmark/gen_image_perturbations.py --dry-run
```

To export the generated images as an additional Hugging Face configuration:

```bash
python benchmark/export_hf_dataset.py --include-perturbed --overwrite
```

### Step 2. (Optional) Regenerate text perturbations

Text perturbations are already stored in `queries/test/metadata.jsonl`. Only rerun this in the source repository if you need to rebuild them:

```bash
python benchmark/gen_text_perturbations.py \
  --gpu 0 --chunk-in chunk_0.json --chunk-out result_0.json
```

---

## Data Layout

```
SnapBench/
├── README.md
├── queries/
│   └── test/
│       ├── metadata.jsonl
│       └── images/          # 1,145 query images
└── gallery/
    └── test/
        ├── metadata.jsonl
        └── images/          # 9,059 unique gallery images
```

The dataset exposes two configurations, both with a `test` split:

- `queries`: clean query images and text, positive and hard-negative gallery IDs, and all text/image perturbation metadata.
- `gallery`: image–caption retrieval candidates. It has 9,085 items backed by 9,059 unique image files because some images have more than one caption.

Important fields:

- Query image: `queries/test/metadata.jsonl` → `file_name` (loaded as `image`)
- Query text: `text`
- Positive gallery items: `positive_gallery_ids`
- Hard negatives: `hard_negative_gallery_ids`
- Gallery image: `gallery/test/metadata.jsonl` → `file_name` (loaded as `image`)
- Gallery caption: `caption`