File size: 5,156 Bytes
2ecd785
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7d13473
2ecd785
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
---
pretty_name: IBCBench
language:
- en
license: other
task_categories:
- other
tags:
- image-retrieval
- multi-image-retrieval
- multimodal-agents
- image-bundle-composition
- visual-history
- yfcc100m
configs:
- config_name: default
  data_files:
  - split: test
    path: queries.jsonl
---

# IBCBench: Image Bundle Composition Benchmark

IBCBench is the benchmark introduced in **Weaving Visual Narratives: Agentic Image Bundle Composition Beyond Atomic Visual Matching**, accepted to the **EMNLP 2026 Main Conference**.

[**GitHub**](https://github.com/LaVieEnRose365/Image-Bundle-Composition) | **Paper link coming soon**

## Overview

Image Bundle Composition (IBC) shifts image retrieval from independently ranking images to dynamically composing a compact, cohesive bundle whose images jointly satisfy relational, temporal, spatial, or narrative constraints. The target bundles are not predefined in the image pool, so systems must reason over non-decomposable bundle-level relevance in a combinatorial search space.

IBCBench contains **667 human-verified queries** evaluated against a pool of **109,467 images from 57 users**. Each ground-truth bundle contains 3--5 images. The benchmark was built using a semi-automated candidate mining and verification pipeline followed by expert human review.

## Dataset Statistics

| Statistic | Value |
|:--|--:|
| Queries | 667 |
| Images | 109,467 |
| Users | 57 |
| Bundle size: 3 images | 162 (24.3%) |
| Bundle size: 4 images | 215 (32.2%) |
| Bundle size: 5 images | 290 (43.5%) |
| Same-location dynamics | 52.5% |
| Cross-location structures | 47.5% |

## Download

Download the complete benchmark from the Hugging Face Hub:

```bash
hf download CyberDancer/IBCBench --repo-type dataset --local-dir IBCBench
```

Then extract the image pool:

```bash
unzip IBCBench/images.zip -d IBCBench
```

Alternatively, the included `download_images.py` script can reconstruct the image pool from `photo_ids/`:

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

## File Structure

```text
IBCBench/
├── queries.jsonl              # 667 IBC queries and ground-truth bundles
├── metadata/
│   └── {user_id}.jsonl        # Photo metadata for each user
├── photo_ids/
│   └── {user_id}.txt          # Photo IDs and storage hashes for each user
├── images.zip                # images/{user_id}/{photo_id}.jpg
└── download_images.py        # Alternative image downloader
```

## Query Format

Each line in `queries.jsonl` is one query and its ground-truth image bundle:

```json
{
  "query_id": "10287726@N02_s18_w0_n3",
  "query": "Find a bundle documenting the same spectator's Olympic viewing experience: ...",
  "image_ids": ["7797992908", "7797994344", "7797998284"]
}
```

| Field | Type | Description |
|:--|:--|:--|
| `query_id` | string | Unique query identifier. Its prefix is the YFCC user ID. |
| `query` | string | Natural-language bundle query in English. |
| `image_ids` | list[string] | Ground-truth set of 3--5 YFCC photo IDs. |

## Photo Metadata Format

Each line in `metadata/{user_id}.jsonl` describes one image:

```json
{
  "photo_id": "4517621778",
  "metadata": {
    "taken_time": "2010-04-10 13:52:57",
    "longitude": -1.239802,
    "latitude": 51.754123,
    "accuracy": 16.0,
    "address": "...",
    "capturedevice": "Panasonic DMC-TZ5"
  }
}
```

The location, address, accuracy, and capture-device fields are optional and may be absent. Album identifiers are intentionally excluded from the benchmark.

Each non-empty line in `photo_ids/{user_id}.txt` has the following tab-separated form:

```text
{photo_id}\t{storage_hash}
```

## Shared Image Pool and Attribution

IBCBench uses the same image pool, metadata, and photo-ID mapping as [DISBench](https://huggingface.co/datasets/RUC-NLPIR/DISBench), which is derived from the public [YFCC100M](https://multimediacommons.wordpress.com/yfcc100m-core-dataset/) collection. IBCBench contributes a separate set of 667 queries and ground-truth bundles for the Image Bundle Composition task.


## Ethical Considerations

The source images were publicly distributed under Creative Commons licenses. IBCBench is intended for research on multimodal retrieval and relational reasoning, particularly user-centric organization of personal photo collections. It is not intended for unauthorized surveillance, identity profiling, or analysis of private third-party collections.

Although the source media is public, the benchmark includes spatiotemporal metadata. Users should handle this information responsibly and follow the terms attached to each original image.

## License

The IBCBench query annotations and BundleWeaver project are released under the [Apache License 2.0](https://github.com/LaVieEnRose365/Image-Bundle-Composition/blob/main/LICENSE). The images retain the individual Creative Commons licenses attached to the corresponding YFCC100M records; downstream users are responsible for checking and following those terms. The included DISBench-derived helper and shared files retain their upstream terms.