README.md

---
license: mit
task_categories:
- object-detection
language:
- en
tags:
- computer-vision
- object-detection
- out-of-distribution
- ood
- evaluation
- robustness
- bdd100k
- pascal-voc
- hallucination-mitigation
size_categories:
- 1K<n<10K
---

# M-Hood Dataset: Out-of-Distribution Evaluation Collection

This dataset collection contains **out-of-distribution (OOD) image datasets** specifically curated for evaluating the robustness of object detection models, particularly those trained to **mitigate hallucination on out-of-distribution data**.

## 🎯 Purpose

These datasets are designed to address limitations in existing OOD benchmarks and enable fine-grained analysis of hallucination suppression. They test how well object detection models perform when encountering images that differ from their training distribution. They are particularly useful for:

-   **Evaluating model robustness** on out-of-distribution data
-   **Testing hallucination mitigation** techniques
-   **Benchmarking domain adaptation** capabilities
-   **Research on robust object detection**

## 📊 Dataset Overview

| Dataset | Images | Description | Domain |
|---------|--------|-------------|---------|
| **far-ood** | 1,000 | Far out-of-distribution images with objects distinctly different from training domains, or backgrounds without recognizable objects. | General OOD |
| **near-ood-bdd** | 1,000 | Near OOD images related to BDD 100K driving domain, visually and semantically similar to training categories. | Autonomous Driving |
| **near-ood-voc** | 1,000 | Near OOD images related to Pascal VOC object classes, visually and semantically similar to training categories. | General Objects |

## 📁 Dataset Structure

```
m-hood-dataset/
├── far-ood/
│   ├── 8a2b026a6c3d5ee2.jpg
│   ├── 5ec941c27b5a6c2f.jpg
│   └── ... (1,000 images)
├── near-ood-bdd/
│   ├── [image files]
│   └── ... (1,000 images)
└── near-ood-voc/
    ├── [image files]
    └── ... (1,000 images)
```

## 🔍 Dataset Details

These datasets were carefully sampled from over 500 diverse categories in OpenImagesV7 to provide challenging and reliable benchmarks for OOD detection in object detection models.

### Far-OOD Dataset
-   **Images**: 1,000 high-quality images
-   **Characteristics**: Images contain objects distinctly different from typical object detection training domains, as well as backgrounds without recognizable objects. This dataset is designed for testing extreme out-of-distribution robustness.

### Near-OOD-BDD Dataset  
-   **Images**: 1,000 high-quality images
-   **Domain**: Related to autonomous driving (BDD 100K-adjacent)
-   **Characteristics**: Images are visually and semantically similar to the training categories of autonomous driving datasets like BDD 100K, presenting a particularly challenging scenario for object detectors.
-   **Use Case**: Testing domain shift robustness in autonomous driving scenarios.

### Near-OOD-VOC Dataset
-   **Images**: 1,000 high-quality images  
-   **Domain**: Related to Pascal VOC object classes
-   **Characteristics**: Images are visually and semantically similar to the training categories of Pascal VOC, presenting a particularly challenging scenario for object detectors.
-   **Use Case**: Testing domain shift robustness for general object detection.

## 🚀 Usage

### Loading with Hugging Face Datasets

```python
from datasets import load_dataset

# Load the entire dataset collection
dataset = load_dataset("HugoHE/m-hood-dataset")

# Access individual subsets
far_ood = dataset["far-ood"]
near_ood_bdd = dataset["near-ood-bdd"] 
near_ood_voc = dataset["near-ood-voc"]
```

### Direct Download

You can also download specific subsets directly:

```python
from huggingface_hub import snapshot_download

# Download specific dataset
snapshot_download(
    repo_id="HugoHE/m-hood-dataset",
    repo_type="dataset",
    local_dir="./datasets",
    allow_patterns="far-ood/*"  # or "near-ood-bdd/*" or "near-ood-voc/*"
)
```

### Evaluation Example

```python
from ultralytics import YOLO
import os
from PIL import Image

# Load your trained model
model = YOLO('path/to/your/model.pt')

# Evaluate on far-ood dataset
far_ood_dir = "path/to/far-ood"
results = []

for img_file in os.listdir(far_ood_dir):
    if img_file.endswith('.jpg'):
        img_path = os.path.join(far_ood_dir, img_file)
        result = model(img_path)
        results.append(result)

# Analyze results for hallucination/false positives
```

## 🔬 Research Applications

This dataset collection is particularly valuable for research in:

-   **Out-of-distribution detection**
-   **Hallucination mitigation in object detection**
-   **Domain adaptation and transfer learning**
-   **Robust computer vision systems**
-   **Autonomous driving perception robustness**
-   **General object detection robustness**

## 📈 Evaluation Metrics

When using these datasets for evaluation, consider these metrics:

-   **False Positive Rate (FPR)**: Rate of hallucinated detections
-   **Confidence Calibration**: How well confidence scores reflect actual accuracy
-   **Detection Consistency**: Consistency of detections across similar OOD images
-   **Domain Shift Sensitivity**: Performance degradation compared to in-distribution data

## 🎯 Related Models

This dataset collection is designed to work with the **M-Hood model collection** available at:
-   **Repository**: [HugoHE/m-hood](https://huggingface.co/HugoHE/m-hood)
-   **Models**: YOLOv10 and Faster R-CNN variants trained on BDD 100K, Pascal VOC, and KITTI
-   **Fine-tuned variants**: Specifically trained to mitigate hallucination on OOD data

## 📄 Citation
If you use our models, datasets, or methodology in your research, please cite our papers.

For the IROS 2025 conference version, which primarily focuses on YOLO models and represents an earlier conference publication, please cite:
```
@inproceedings{he2025mitigating,
      title={Mitigating Hallucinations in YOLO-based Object Detection Models: A Revisit to Out-of-Distribution Detection},
      author={Weicheng He and Changshun Wu and Chih-Hong Cheng and Xiaowei Huang and Saddek Bensalem},
      booktitle={Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
      year={2025},
      note={Accepted to IROS 2025},
      eprint={2503.07330v2},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.07330v2}
}
```
For the journal version, which expands the methodology to Faster-RCNN and RT-DETR, includes an automated data curation pipeline, and provides an in-depth analysis of the approach, please cite:
```
@inproceedings{wu2025revisiting,
      title={Revisiting Out-of-Distribution Detection in Real-time Object Detection: From Benchmark Pitfalls to a New Mitigation Paradigm}, 
      author={Changshun Wu and Weicheng He and Chih-Hong Cheng and Xiaowei Huang and Saddek Bensalem},
      year={2025},
      eprint={2503.07330v3},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.07330v3}, 
}
```
*Note: These datasets were constructed using an automated data curation pipeline as part of the M-Hood project, originally described at [https://gricad-gitlab.univ-grenoble-alpes.fr/dnn-safety/m-hood](https://gricad-gitlab.univ-grenoble-alpes.fr/dnn-safety/m-hood).*

## 📜 License

This dataset collection is released under the MIT License.

## 🏷️ Keywords

Out-of-Distribution, OOD, Object Detection, Computer Vision, Robustness Evaluation, Hallucination Mitigation, BDD 100K, Pascal VOC, Domain Adaptation, Model Evaluation.