YOLO11 Normalization

[bastithebast]

Hi,
It is stated that for the image normalization predefined mean and standard deviation values are used (see: https://docs.ultralytics.com/guides/preprocessing_annotated_data/#normalizing-pixel-values).
How exactly are these values predefined? Is it just general values calculated from a lot of different datasets or is the training dataset analyzed before the actual training and the mean and std calculated for this specific dataset?

And where can I find this information? I already tried to find it in the YOLO11 paper, but without success. But maybe I missed it.

Thanks a lot for your help!

[UltralyticsAssistant]

👋 Hello @bastithebast, thank you for your interest in Ultralytics 🚀! We recommend a visit to the Docs for new users, where you'll find a wealth of resources including guides on preprocessing annotated data that may address parts of your question.

If this is a ❓ question related to custom training, please provide further details if applicable, such as your dataset, code snippets, and specific preprocessing methods used so far. This context will help both us and the community better understand your query.

If this is a 🐛 Bug Report or involves an issue in your workflow, please provide a minimum reproducible example (MRE) to assist us in debugging effectively.

Related Points of Interest

1. The details on normalization values, as mentioned in the Docs, could be useful to explore.
2. For a deeper dive into YOLO11 settings, make sure your local ultralytics package is updated to the latest version. You can upgrade with:

   pip install -U ultralytics

Additional Resources

Join the Ultralytics community to connect with others:

• For real-time discussions, hop into our Discord community 🎧.
• Explore broader questions or share knowledge on Discourse or Subreddit.

Validated Environments

Ensure your setup aligns with one of our validated environments:

• Run in a notebook with a free GPU:
• Configure with GCP Quickstart Guide, AWS Quickstart Guide, or the pre-built Docker image

👉 Rest assured, this is an automated response, but an Ultralytics engineer will review and address your question and concerns soon. Thank you for being an active member of the YOLO community! 🌟

[glenn-jocher]

@bastithebast thanks for your question! YOLO11 uses fixed normalization values ([0.0, 0.0, 0.0] for mean and [255.0, 255.0, 255.0] for std) that scale pixel values to 0-1 range by default, rather than dataset-specific calculations. These conventions are consistent across our implementations to simplify training pipelines. You can find the normalization logic in the dataloaders at https://github.com/ultralytics/ultralytics/blob/main/ultralytics/data/dataloaders/v5augmentations.py.

[glenn-jocher]

@bastithebast yes, you can customize normalization values by modifying the data loader's mean and std parameters in your training configuration or by creating a custom data augmentation pipeline that applies your dataset-specific normalization before training. This requires overriding the default preprocessing transforms in the YOLO data pipeline.

[bastithebast]

Thanks for your help!

[glenn-jocher]

You can customize normalization values by modifying the mean and std parameters in your training configuration or data augmentation pipeline. For implementation details, see the classify_transforms function in our docs. Happy training! 🚀

[bastithebast]

I am sorry, I can see how to use this function before making a prediction on individual images, but I am still not quite sure how to use classify_transforms on a whole dataset before training.

I found the following code in another discussion from a year ago:

from ultralytics import YOLO, Augment
import albumentations as A

# Define your custom augmentations
custom_augmentations = Augment(
    A.Compose([
        A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2, p=0.5),
        A.RandomGamma(gamma_limit=(80, 120), p=0.5),
        A.ImageCompression(quality_lower=75, quality_upper=99, p=0.5)
    ])
)

# Create a YOLO model and pass the custom augmentations
model = YOLO('yolov8n.yaml', augment=custom_augmentations)

# Train the model
model.train(data='coco128.yaml', epochs=3)

So I tried the following:

from ultralytics import YOLO
from ultralytics.data.augment import classify_transforms

# bad augmentation to test the effect
new_mean = (0.9, 0.9, 0.9)
new_std = (0.1, 0.1, 0.1)

imgsz = 640

custom_augmentations = classify_transforms(size=imgsz,mean=new_mean, std=new_std )

model = YOLO("yolo11m.pt", augment=custom_augmentations)
model.train(data=data.yaml, epochs=100, imgsz=imgsz, batch=16, project='path/to/project', name='name', device='0,1')

But I get: TypeError: init() got an unexpected keyword argument 'augment'

I also, somewhat desperately, tried to include this in my data.yaml file, but with no suprise without any effect:

augmentations:
  size: 640
  mean: (0.9, 0.9, 0.9)
  std: (0.1, 0.1, 0.1)

I would be very thankful if you could tell me what I should do instead!

[glenn-jocher]

@bastithebast to apply custom normalization during YOLO11 training, modify your data.yaml with these parameters under the augment section:

augment:
  mean: [0.9, 0.9, 0.9]
  std: [0.1, 0.1, 0.1]

The augment parameter in YOLO initialization is deprecated. For implementation details, see our updated data augmentation documentation. Let us know if you encounter any other issues!