ImageRetrieval

Installation

To use the ImageRetrieval package, you can follow these installation steps:

1. Clone the repository:

```bash
git clone https://github.com/your-username/your-repo.git
cd your-repo
```

2. Install the ImageRetrieval package itself:

```bash
pip install .
```

This will automatically install packages

3. You're all set! Now lets look at results from our approach

Configuration	Model	Loss Functions	Feature Dim	Augmentation	ROC-AUC
0	Resnet50	-	-	-	0.75
1	Resnet50	CenterLoss + CrossEntropy	512	False	0.815
2	Resnet50	CenterLoss + CrossEntropy + ArcFaceLoss	512	True	0.84
3	Resnet50	CenterLoss + CrossEntropy + ArcFaceLoss	4096	True	0.85
4	Resnet101	CenterLoss + CrossEntropy + ArcFaceLoss	4096	True	0.87
5	Hybrid_CNN_Transformer	CenterLoss + CrossEntropy + ArcFaceLoss	8192	True	0.9

Explanation

In this project, a range of model configurations were explored to improve image retrieval performance, with ROC-AUC scores used as the evaluation metric.

• Resnet50 Inference (Frozen Layers): The initial model achieved a ROC-AUC of 0.75.
• Resnet50 with CenterLoss + CrossEntropy: Introducing the CenterLoss and CrossEntropy loss functions along with a feature dimension of 512 improved the performance to a ROC-AUC of 0.815.
• Resnet50 with CenterLoss + CrossEntropy + ArcFaceLoss: Adding ArcFaceLoss and enabling augmentation further improved performance, yielding a ROC-AUC of 0.84.
• Resnet50 with CenterLoss + CrossEntropy + ArcFaceLoss (FeatureDim=4096): Increasing the feature dimension to 4096 while retaining augmentation led to a ROC-AUC of 0.85.
• Resnet101 with CenterLoss + CrossEntropy + ArcFaceLoss (FeatureDim=4096): Using a Resnet101 model with augmented feature dimensions achieved a ROC-AUC of 0.87.
• Hybrid CNN-Transformer (Resnet101 + SwinTransformer): Combining a Resnet101 CNN with a SwinTransformer and augmenting the feature dimension to 8192 resulted in an exceptional ROC-AUC of 0.9.

These configurations showcase the impact of diverse strategies on the ROC-AUC score, with augmentation, feature dimension adjustments, and hybrid architectures playing significant roles in performance enhancements. The success of the hybrid model underscores the benefits of integrating transformer-based features alongside CNNs.

---

Feel free to copy and paste this content into your GitHub README as desired.

t-SNE Visualization of Classes

The image above showcases the t-SNE visualization of classes, offering insights into the distribution of your data