Confusion matrix

[glenn-jocher]

This PR implements confusion matrices using code from:
https://github.com/kaanakan/object_detection_confusion_matrix

VOC Confusion Matrix

python test.py --weights last.pt --data voc.yaml --img 512

COCO Confusion Matrix

python test.py --weights last.pt --data coco.yaml --img 640

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced testing capabilities in YOLOv5 with a new confusion matrix feature and improved results logging.

📊 Key Changes

• Added seaborn and pandas to requirements.txt, previously commented out.
• Introduced ConfusionMatrix class in utils/metrics.py to calculate and plot confusion matrices.
• Updated test.py to create a confusion matrix instance and process batches to update this matrix.
• Modified test.py and train.py to plot and save the confusion matrix during testing and training.
• Adjusted WandB logging to include confusion matrix images where applicable.
• Changed requirement management by moving pycocotools out of the commentary section for COCO mAP computation.

🎯 Purpose & Impact

• These additions will provide users with a detailed visual understanding of model performance with confusion matrices.
• Enhanced analytical capabilities will aid in identifying classification errors more systematically.
• By introducing key data analysis libraries, YOLOv5 construes to be not just a training toolkit but also a model analysis suite.
• Users leveraging WandB for experiment tracking will now have a richer set of visualizations automatically logged.
• Ultimately, a better grasp of a model's predictions across different classes may hasten iterations on model improvement.