README.md

---
language:
  - en
license: unknown
pretty_name: PittImageVideoAdsDataset
tags:
  - advertising
  - image-advertisements
  - video-advertisements
  - visual-understanding
  - graphic-design
annotations_creators:
  - crowdsourced
language_creators:
  - found
size_categories:
  - 10K<n<100K
source_datasets:
  - original
task_categories:
  - image-classification
configs:
  - config_name: image_ads
    data_files:
      - split: train
        path: image_ads/train-*
  - config_name: video_ads
    data_files:
      - split: train
        path: video_ads/train-*
---

# Dataset Card for PittImageVideoAdsDataset

[![CI](https://github.com/creative-graphic-design/huggingface-datasets/actions/workflows/ci.yaml/badge.svg)](https://github.com/creative-graphic-design/huggingface-datasets/actions/workflows/ci.yaml)
[![Sync HF](https://github.com/creative-graphic-design/huggingface-datasets/actions/workflows/push_to_hub.yaml/badge.svg)](https://github.com/creative-graphic-design/huggingface-datasets/actions/workflows/push_to_hub.yaml)

## Dataset Description

- **Homepage:** https://people.cs.pitt.edu/~kovashka/ads/
- **Repository:** https://github.com/creative-graphic-design/huggingface-datasets/tree/main/datasets/PittImageVideoAdsDataset
- **Hugging Face Dataset:** https://huggingface.co/datasets/creative-graphic-design/PittImageVideoAdsDataset
- **Paper (arXiv):** https://arxiv.org/abs/1707.03067
- **Paper (CVPR 2017):** https://openaccess.thecvf.com/content_cvpr_2017/html/Hussain_Automatic_Understanding_of_CVPR_2017_paper.html
- **Official Image Annotations:** https://people.cs.pitt.edu/~kovashka/ads/annotations_images.zip
- **Official Video Annotations:** https://people.cs.pitt.edu/~kovashka/ads/annotations_videos.zip

### Dataset Summary

PittImageVideoAdsDataset is the image and video advertisement dataset released with *Automatic Understanding of Image and Video Advertisements*. The paper reports 64,832 image advertisements and 3,477 YouTube advertisement videos, with human annotations for topics, sentiments, slogans, persuasive strategies, symbolic references, and action/reason Q/A. This Hugging Face version exposes the public annotation rows that can be reconstructed from the official released files.

This loader exposes two configurations:

- `image_ads`: annotated image advertisements. The `image` column is a `datasets.Image` feature loaded from the official Google Storage image ZIP files, and `image_url` preserves the corresponding Pitt source URL as text.
- `video_ads`: YouTube video IDs and URLs with raw annotator responses and cleaned majority-vote labels. Video files are not downloaded.

### Supported Tasks and Leaderboards

The dataset supports advertisement understanding tasks such as topic recognition, sentiment recognition, persuasive strategy analysis, symbol detection, and action/reason response modeling. No active public leaderboard is bundled with this Hugging Face dataset.

### Languages

The annotation text and Q/A responses are in English (`en`).

## Dataset Structure

### Data Instances

PittImageVideoAdsDataset provides `image_ads` and `video_ads` configurations. Load one configuration by passing its name:

```python
import datasets as ds

dataset = ds.load_dataset(
    "creative-graphic-design/PittImageVideoAdsDataset",
    name="image_ads",
)
```

`image_ads` example:

```json
{
  "id": "image-00000",
  "source_path": "0/10000.jpg",
  "image": "<remote image>",
  "image_url": "https://people.cs.pitt.edu/~mzhang/image_ads/0/10000.jpg",
  "qa_action": ["..."],
  "qa_reason": ["..."],
  "qa_combined_action_reason": [],
  "slogans": [],
  "topics": ["9", "9", "9"],
  "sentiments": [["12"], ["14"]],
  "strategies": [],
  "symbols": []
}
```

`video_ads` example:

```json
{
  "id": "video-00000",
  "youtube_id": "KONL05sae4E",
  "youtube_url": "https://www.youtube.com/watch?v=KONL05sae4E",
  "raw_qa_action": ["Buy and wear this company's sports shoes and clothes."],
  "raw_qa_reason": ["..."],
  "raw_topics": ["media", "media"],
  "raw_sentiments": ["active"],
  "raw_funny": [0.0, 0.0, 0.0, 0.0, 0.0],
  "raw_exciting": [1.0, 1.0, 1.0, 1.0, 1.0],
  "raw_language": ["1", "1", "1", "1", "1"],
  "clean_topic": 27,
  "clean_sentiment": 12,
  "clean_funny": 0.0,
  "clean_exciting": 1.0,
  "clean_language": "1",
  "clean_effective": 1.0
}
```

### Data Fields

`image_ads` fields:

- `id` (`string`): Stable row identifier generated by this loader.
- `source_path` (`string`): Original relative image path from the annotation files.
- `image` (`Image`): Advertisement image loaded from the official image ZIP files.
- `image_url` (`string`): Pitt source image URL derived from `source_path`.
- `qa_action`, `qa_reason`, `qa_combined_action_reason`, `slogans`, `topics` (`list[string]`): Free-form or class-id annotations.
- `sentiments`, `strategies` (`list[list[string]]`): Per-annotator class-id selections.
- `symbols` (`list`): Symbolic reference boxes with `x1`, `y1`, `x2`, `y2`, and `label`.

`video_ads` fields:

- `id`, `youtube_id`, `youtube_url` (`string`): Stable row identifier and YouTube reference.
- `raw_qa_action`, `raw_qa_reason`, `raw_topics`, `raw_sentiments`, `raw_language` (`list[string]`): Raw annotator responses or labels.
- `raw_funny`, `raw_exciting` (`list[float]`): Raw per-annotator binary selections.
- `clean_topic`, `clean_sentiment` (`int32`): Cleaned majority-vote class IDs.
- `clean_funny`, `clean_exciting`, `clean_effective` (`float32`): Cleaned binary labels.
- `clean_language` (`string`): Cleaned language label. Some videos do not have this field upstream and use an empty string.

### Data Splits

Both configurations expose one `train` split because the upstream release is an annotation corpus and does not define train, validation, or test partitions. The `train` name is a Hugging Face packaging convention here; it is not an official training partition from the paper.

| Config | Split | Rows |
| --- | --- | ---: |
| `image_ads` | train | 64,454 annotated image paths |
| `video_ads` | train | 3,477 videos |

The row counts line up with the public files as follows:

| Source count | Images | Videos |
| --- | ---: | ---: |
| Paper-reported dataset size | 64,832 | 3,477 |
| Rows emitted by this loader | 64,454 | 3,477 |

The image count is lower than the paper-reported total because this loader emits the union of image paths found in the public annotation JSON files. The remaining paper-reported images are not present as annotated rows in those JSON files, so the loader does not synthesize empty examples for them.

## Dataset Creation

### Curation Rationale

The dataset was created to study automatic understanding of advertisements beyond object recognition, including what action an ad encourages, why a viewer should take that action, topic, sentiment, and persuasive visual strategies.

### Source Data

The original project collected image advertisements and YouTube advertisement videos. The official project page provides annotation ZIP files, video IDs, Pitt image URLs, and Google Storage image ZIP files. This loader downloads the official annotation ZIP files and image ZIP files, then derives Pitt image URLs directly from the annotation `source_path` values.

### Annotations

The annotations were collected from human annotators, including free-form Q/A and slogans, multiple-choice topics and sentiments, symbolic reference boxes, and cleaned majority-vote labels for video-level attributes.

### Personal and Sensitive Information

Advertisements may contain people, brands, public figures, health claims, political content, or other persuasive and sensitive themes. Users should inspect examples and labels before applying the dataset to downstream systems.

## Considerations for Using the Data

### Social Impact of Dataset

This dataset can help evaluate and improve models that reason about persuasive visual media. Such models may also be used for ad targeting, content moderation, or influence analysis, so downstream use should consider privacy, bias, and manipulation risks.

### Discussion of Biases

The dataset reflects the distribution of advertisements collected by the original authors and the judgments of crowd annotators. Topics, sentiments, and persuasive strategies may encode cultural assumptions from the source ads and annotator pool.

### Other Known Limitations

The `image` column embeds decoded image data when this dataset is published to the Hub. The `image_url` field remains a source reference and may be less stable from automated environments. The video configuration contains YouTube IDs and URLs, not video files, so availability may change if YouTube videos are removed or restricted.

## Additional Information

### Dataset Curators

The dataset was created by the authors of *Automatic Understanding of Image and Video Advertisements*. This Hugging Face dataset loader was added in the `creative-graphic-design/huggingface-datasets` repository.

### Licensing Information

A dataset-specific license was not found on the official project page at the time this loader was created. This dataset card marks the dataset content license as `unknown`.

### Citation Information

```bibtex
@inproceedings{hussain2017automatic,
  title={Automatic Understanding of Image and Video Advertisements},
  author={Hussain, Zaeem and Zhang, Mingda and Zhang, Xiaozhong and Ye, Keren and Thomas, Christopher and Agha, Zuha and Ong, Nathan and Kovashka, Adriana},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={1705--1715},
  year={2017}
}
```

### Contributions

Thanks to Zaeem Hussain, Mingda Zhang, Xiaozhong Zhang, Keren Ye, Christopher Thomas, Zuha Agha, Nathan Ong, Adriana Kovashka, and the University of Pittsburgh project team for creating and releasing this dataset.