AISystem-2402

This repository provides implementations of both few-shot and zero-shot anomaly detection using CLIP embeddings.

Project Structure

AISystem-2402
├── fewshot-clip/
├── zeroshot-clip/
├── .gitignore
├── LICENSE
└── README.md

Dataset Structure

Training Dataset

└──── ./train
    ├──── {class_name}/
    |   ├──── 0000.jpg
    |   ├──── ...
    |   └──── 0000.jpg
    ├──── {class_name}/
    |   ├──── 0000.jpg
    |   ├──── ...
    |   └──── 0000.jpg

Test (Validation) Dataset

└──── ./test
    ├──── anomaly/
    |   ├──── 0000.jpg
    |   ├──── ...
    |   └──── 0000.jpg
    ├──── normal/
    |   ├──── 0000.jpg
    |   ├──── ...
    |   └──── 0000.jpg
    

The dataset for anomaly detection is organized to facilitate both few-shot and zero-shot approaches with CLIP embeddings.

Dataset Download

Dataset can be downloaded here.(Last updated on November 24, 2024)

Dataset Organization

For Preliminary Competition (for personal practice)

• Training Data: 6 classes, 5 normal images per class.
• Validation Data: 6 classes, 200 anomaly and 200 normal images for each class.

For Main Competition (for project)

• Training Data: 10 classes (excluding the 6 classes from the preliminary competition), 5 normal images per class.
• Validation Data: No validation dataset will be given.

Each class within the anomaly dataset contains at least one example of an anomaly, such as dots, cuts, or other class-specific defects like below. This design encourages exploration of anomaly detection within constrained data conditions.

Getting Started

1. Clone the Repository

git clone https://github.com/PiLab-CAU/AISystem-2402.git

2. Set up the Conda Environment

conda create -n cauclip python=3.xx

The Recommended Python version is 3.8 or above 😉

conda activate cauclip

3. Navigate to the Cloned Folder

cd AISystem-2402

4. Install Dependencies

pip install -r requirements.txt

Once the dependencies are installed, we are ready to detect anomalies! 😀

Run Sample Code

Fewshot-CLIP

To run the sample code for Few-Shot CLIP Embedding, execute:

python fewshot-clip/main.py

Zeroshot-CLIP

To run the sample code for Zero-Shot CLIP Embedding, execute:

python zeroshot-clip/main.py

Performance Summary

The Performance of sample codes is shown in the tables below.

Few-Shot CLIP

Metric	Value
Train Loss (Epoch 1)	0.7279
Train Accuracy (Epoch 1)	52.86%
Test Accuracy (Epoch 1)	63.08%
Best Model Path	saved_models/best_model.pth (Accuracy: 63.08%)
Overall Test Accuracy	32.31%
Total Time Taken	18.42 seconds
Average Time per Image	0.2834 seconds

Inference Details

Class	Correct	Total	Accuracy	Normal Similarity	Anomaly Similarity
Normal	21	21	100.00%	-	-
Anomaly	0	44	0.00%	-	-

---

Zero-Shot CLIP

Metric	Value
Total Images	65
Correct Predictions	35
Overall Accuracy	53.85%

Class-Specific Performance

Class	Total	Correct	Incorrect	Accuracy	Avg Anomaly Score	Avg Normal Similarity	Avg Anomaly Similarity
Normal	21	21	0	100.00%	0.288	0.921	0.634
Anomaly	44	14	30	31.82%	0.234	0.862	0.627

Detailed Metrics

Metric	Value
Precision	100.00%
Recall	31.82%
F1 Score	48.28%

Metrics saved in ./results/metrics_{datetime}_{time}.json. Results can be found in the ./results directory.

Further Information

For additional details on each module, check out the specific README files:

• Few-Shot CLIP

• Zero-Shot CLIP

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contact

For questions or further information, please contact leessoit@gmail.com or use the issue tab to report any problems or suggestions.