Problem with augmentation

[NS-Nik]

Hello YOLO community!

I have trained the YOLO11 medum model to detect ocean eddies from satellite SAR images. I have two classes based on eddy rotation: cyclones and anticyclones:

So far so good, but I have limited number of images and I am trying to implement augmentation to increase my training sample and metrics.
However, I cannot use the "flip" augmentation, since image flipping leads to the class change. For example, If I flip the image with anticyclonic eddy it becomes an eddy with cyclonic rotation.

My question is the following: I want to flip images in the training dataset prior to the model training and change the classes in the corresponding labels and then add this data to the training dataset. Is it a proper way to increase my training sample?

Or maybe there are some other ways to tackle this issue?

[UltralyticsAssistant]

👋 Hello @NS-Nik, thank you for sharing your question with the Ultralytics community 🚀! It sounds like you're working on a fascinating application of YOLO for ocean eddy detection 🌊. Your description of the augmentation challenge involving flipping and class changes is quite insightful.

If this is a custom training ❓ question, providing as much detail as possible—including any specific examples of your current dataset, training scripts, and results so far—can help us understand the context better. Also, ensuring you're following best practices for training (such as proper dataset preparation and augmentation techniques) can greatly improve your results.

Alternatively, if you’re encountering issues with augmentation behavior in the framework that might indicate a 🐛 bug, we recommend including a minimum reproducible example (MRE) to help pinpoint the issue more effectively.

Make sure that you’re using the latest version of the ultralytics package by upgrading it in your Python environment to ensure any potential issues are not due to an outdated version. For your question about augmentation, manually flipping your images and updating the labels accordingly, as you described, is indeed a reasonable approach if done carefully. However, experimenting with other augmentation techniques (like rotation, scaling, or color adjustments) might also yield good results while avoiding the class-label flipping issue.

This is an automated response to provide you with an overview of potential resources and actions, but don’t worry—a member of our Ultralytics team will review your discussion soon to assist further. Keep up the great work, and thank you for contributing to the Ultralytics community! 🚀

[glenn-jocher]

@NS-Nik thank you for your question! Flipping the dataset and updating the labels accordingly is a valid approach to augment your training data while addressing the class change issue. Alternatively, you can consider using augmentation techniques like rotation, cropping, scaling, or color adjustments that do not alter the fundamental nature of the classes. Tools like Albumentations can help you implement these transformations effectively. For more details on augmentation techniques, refer to our guide.

[NS-Nik]

Thank you for the answer!
I implement all the mentioned augmentation techniques besides flipping, due to the class change.
Now I just add the flip by flipping the dataset prior to the training.

[glenn-jocher]

@NS-Nik you're welcome! Adding the flipped dataset prior to training is a great approach. Ensure the class labels are updated accurately for the flipped images, and you should see improved model performance.

[NS-Nik]

@glenn-jocher
Btw. Should I add the images from the augmented dataset to the validation batch? Or just use it for training?

[glenn-jocher]

@NS-Nik you should only use the augmented dataset for training and keep the validation set unaltered to accurately evaluate your model's performance on original, unseen data. For details, refer to this guide.