README.md

---
license: cc-by-4.0
task_categories:
- object-detection
- image-classification
- image-to-image
- text-to-image
language:
- en
pretty_name: >-
  AI-generated Dataset of Outpainted Vehicles for Eye-level Classification and
  Localization
size_categories:
- 10K<n<100K
viewer: true
tags:
- transportation
- autonomous-driving
- traffic-surveillance
- vehicle-image
---
# AIDOVECL: AI-generated Dataset of Outpainted Vehicles for Eye-level Classification and Localization

<!-- Provide a quick summary of the dataset. -->

We introduce an annotated AI-generated dataset of eye-level vehicle images using outpainting, offering versatile generation of diverse vehicle classes in varied contexts with pretrained models.


<h1 style="color: blue;">Citation Notice</h1>

Please ensure that all publications and presentations using this data reference the following paper:

[Kazemi, A., Fatima, Q., Kindratenko, V., & Tessum, C. (2024). AIDOVECL: AI-generated Dataset of Outpainted Vehicles for Eye-level Classification and Localization. arXiv preprint arXiv:2410.24116.
](https://arxiv.org/abs/2410.24116)


## Dataset Details

### Dataset Description

<!-- Provide a longer summary of what this dataset is. -->

Our dataset consists of AI-generated, eye-level vehicle images created by detecting and cropping vehicles from manually chosen and classified seed images. These vehicles are then outpainted onto larger canvases to replicate diverse real-world scenarios. The resulting images are meticulously annotated automatically, offering high-quality ground truth data. Through the use of sophisticated outpainting techniques and stringent image quality assessments, we ensure exceptional visual fidelity and contextual accuracy. Moreover, the efficiency and utility of the dataset are demonstrated through benchmark experiments presented in the accompanying paper.


- **Curated by:** Amir Kazemi, Qurat ul ain Fatima, Volodymyr Kindratenko, Christopher W Tessum.
- **Funded by:** The authors acknowledge support from the University of Illinois National Center for Supercomputing Applications (NCSA) Faculty Fellowship Program, the Zhejiang University and University of Illinois Dynamic Research Enterprise for Multidisciplinary Engineering Sciences (DREMES), and the University of Illinois Discovery Partners Institute (DPI).
This work utilizes resources supported by the National Science Foundation’s (NSF) Major Research Instrumentation program, grant #1725729, as well as the University of Illinois at Urbana-Champaign.
- **License:** Creative Commons Attribution 4.0 International

### Dataset Sources

<!-- Provide the basic links for the dataset. -->

- **Repository:** https://huggingface.co/datasets/amir-kazemi/aidovecl/tree/main
- **Paper:** https://arxiv.org/abs/2410.24116
- **Demo:** https://github.com/amir-kazemi/aidovecl

## Uses

<!-- Address questions around how the dataset is intended to be used. -->

### Direct Use

<!-- This section describes suitable use cases for the dataset. -->
This curated dataset combines real, manually classified vehicle images with AI-generated outpainted variants, offering a diverse yet structured resource for computer vision and mobility-related applications. The images maintain contextual plausibility and consistent labeling, making the dataset suitable for training, evaluation, and augmentation workflows. In autonomous driving, it supports vehicle detection and recognition by exposing models to varied environments and viewpoints. The inclusion of fine-grained classes—reflecting differences in vehicle size, passenger capacity, and utility—enables more detailed analysis in urban planning and transportation modeling, such as estimating fleet composition, mode share, and curb usage. For environmental research, the dataset facilitates modeling of vehicle-type-specific emissions and spatial distribution patterns. By combining real-world fidelity with scalable synthetic augmentation, it contributes to more reproducible and generalizable pipelines across both technical and policy-oriented domains.

### Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the dataset will not work well for. -->
While the dataset improves access to diverse and well-annotated vehicle imagery, it is not intended for all uses. Malicious or unethical applications—such as unauthorized surveillance or activities that infringe on privacy—are strictly prohibited. The dataset is most suitable for domains like autonomous driving, urban planning, and environmental monitoring, where the vehicle-level annotations and scene contexts are relevant. Applying it in unrelated fields, especially those misaligned with its structure or labeling scheme, may lead to poor performance or misleading outcomes.

## Dataset Structure

<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->

The root directory contains the augmented dataset, curated by manually excluding rare generation failures to ensure aesthetic quality and realism, and organized into train, val, and test splits. In contrast, the datasets.zip archive contains the benchmarking datasets used in the paper, structured for reproducibility. It includes real_raw, which holds the original manually classified images grouped by vehicle type; real, a YOLO-formatted selection of these images with corresponding labels and a configuration file (real.yaml); and augmented, which extends the real set with AI-generated outpainted images and labels, accompanied by augmented.yaml.

### Vehicle Classification and Subcategories

| ID | Vehicle Class | Subcategories |
|----|---------------|----------------------------------------------------------------|
| 0  | COUPE         | Coupe, Convertible, Cabriolet, or other two-door passenger cars |
| 1  | SEDAN         | Sedan, or other four-door passenger cars |
| 2  | SUV           | SUV or Crossover |
| 3  | MINIVAN       | Minivan or Wagon |
| 4  | MINIBUS       | Minibus, Shuttles, or large passenger vans |
| 5  | BUS           | City, Coach, Double-Decker, Articulated, or School Bus |
| 6  | VAN           | Work, Camper, or Conversion Van |
| 7  | PICKUP        | Regular, Crew Cab, or Extended Cab Pickup Truck |
| 8  | TRUCK         | Single Unit, Trailer, Articulated, Dump, Tanker, or Mixer Truck |

### Split Summary (image/object counts per class)

| split | background | bus  | coupe | minibus | minivan | pickup | sedan | suv  | truck | van | TOTAL |
|-------|------------|------|-------|---------|---------|--------|-------|------|-------|-----|-------|
| train | 784        | 1715 | 1209  | 330     | 350     | 356    | 1642  | 726  | 1071  | 269 | 8452  |
| val   | 98         | 214  | 152   | 40      | 45      | 44     | 205   | 91   | 134   | 32  | 1055  |
| test  | 98         | 213  | 150   | 41      | 43      | 44     | 205   | 91   | 133   | 33  | 1051  |


## Dataset Creation

### Curation Rationale

<!-- Motivation for the creation of this dataset. -->

Publicly available vehicle recognition datasets have driven progress in the field but often fall short in key areas, including lacking contextual richness, offering limited vehicle classifications, and presenting suboptimal viewing angles. These shortcomings underscore the necessity for customizable vehicle datasets tailored to meet the specific requirements of contemporary vehicle recognition applications.

### Source Data

<!-- This section describes the source data (e.g. news text and headlines, social media posts, translated sentences, ...). -->

#### Data Collection and Processing

<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->

The data collection and processing procedures are documented in detail in the manuscript.

#### Who are the source data producers?

<!-- This section describes the people or systems who originally created the data. It should also include self-reported demographic or identity information for the source data creators if this information is available. -->

The origin of the source data is described in detail in the manuscript.

### Annotations

<!-- If the dataset contains annotations which are not part of the initial data collection, use this section to describe them. -->

#### Annotation process

<!-- This section describes the annotation process such as annotation tools used in the process, the amount of data annotated, annotation guidelines provided to the annotators, interannotator statistics, annotation validation, etc. -->

All annotations in this study were generated automatically using the pipeline described in the manuscript.

#### Who are the annotators?

<!-- This section describes the people or systems who created the annotations. -->

No manual annotators were involved.

#### Personal and Sensitive Information

<!-- State whether the dataset contains data that might be considered personal, sensitive, or private (e.g., data that reveals addresses, uniquely identifiable names or aliases, racial or ethnic origins, sexual orientations, religious beliefs, political opinions, financial or health data, etc.). If efforts were made to anonymize the data, describe the anonymization process. -->

To the best of our knowledge, the dataset does not contain information that is likely to be personal or sensitive, nor is any content intended to identify individuals.

## Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->

Our dataset currently cannot outpaint multiple vehicles in a single image due to the limitations of the inpainting model, which struggles with generating coherent scenes with multiple vehicles. This also stems from the challenge of selecting seed images that naturally correlate. Consequently, our dataset lacks rich occlusion scenarios crucial for training robust detection systems. Additionally, our reliance on pretrained detection and inpainting models means that objects not included in the original model training may be missed or appear unrealistic. We also face a shortage of seed images, especially for certain vehicle classes like vans, leading to class imbalances. Addressing these issues is crucial for improving our dataset's effectiveness and the accuracy of detection models.

### Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->

As future work, we recommend exploring more advanced generative methods or complementary augmentations to better capture multi-vehicle and occlusion scenarios, expanding seed image collections to address class imbalance, and fine-tuning pretrained models with domain-specific data for broader coverage. These efforts should be paired with systematic quality checks, human oversight, and transparent documentation to ensure realism, mitigate bias, and guide appropriate dataset use.

## Citation

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**

```bibtex
@article{kazemi2024aidovecl,
  title={AIDOVECL: AI-generated Dataset of Outpainted Vehicles for Eye-level Classification and Localization},
  author={Kazemi, Amir and Kindratenko, Volodymyr and Tessum, Christopher and others},
  journal={arXiv preprint arXiv:2410.24116},
  year={2024}
}
```

**APA:**
```bibtex
Kazemi, A., Kindratenko, V., & Tessum, C. (2024). AIDOVECL: AI-generated Dataset of Outpainted Vehicles for Eye-level Classification and Localization. arXiv preprint arXiv:2410.24116.
```
## Dataset Card Authors

Amir Kazemi

## Dataset Card Contact

[email protected]

[email protected]