Update README.md

README.md

@@ -24,20 +24,20 @@ This dataset collection contains **out-of-distribution (OOD) image datasets** sp
 
 ## 🎯 Purpose
 
-These datasets are designed to test how well object detection models perform when encountering images that differ from their training distribution. They are particularly useful for:
+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**
+-   **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 | Size | Description | Domain |
-|---------|--------|------|-------------|---------|
-| **far-ood** | 1,000 | 278MB | Far out-of-distribution images significantly different from training domains | General OOD |
-| **near-ood-bdd** | 1,010 | 337MB | Near OOD images related to BDD 100K driving domain | Autonomous Driving |
-| **near-ood-voc** | 1,020 | 318MB | Near OOD images related to Pascal VOC object classes | General Objects |
+| 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
 
@@ -49,33 +49,31 @@ m-hood-dataset/
 │   └── ... (1,000 images)
 ├── near-ood-bdd/
 │   ├── [image files]
-│   └── ... (1,010 images)
+│   └── ... (1,000 images)
 └── near-ood-voc/
     ├── [image files]
-    └── ... (1,020 images)
+    └── ... (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
-- **Size**: 278MB
-- **Characteristics**: Images significantly different from typical object detection training domains
-- **Use Case**: Testing extreme out-of-distribution robustness
+-   **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**: 1,010 high-quality images
-- **Size**: 337MB
-- **Domain**: Related to autonomous driving (BDD 100K-adjacent)
-- **Characteristics**: Images similar to but distinct from BDD 100K training distribution
-- **Use Case**: Testing domain shift robustness in autonomous driving scenarios
+-   **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**: 1,020 high-quality images  
-- **Size**: 318MB
-- **Domain**: Related to Pascal VOC object classes
-- **Characteristics**: Images similar to but distinct from Pascal VOC training distribution
-- **Use Case**: Testing domain shift robustness for general object detection
+-   **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
 
@@ -136,28 +134,28 @@ for img_file in os.listdir(far_ood_dir):
 
 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**
+-   **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
+-   **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
+-   **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
 
@@ -171,6 +169,7 @@ If you use this dataset collection in your research, please cite:
   howpublished={\url{https://huggingface.co/datasets/HugoHE/m-hood-dataset}}
 }
 ```
+*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