Correct image transforms for Yolo-NAS

[VadimShabashov]

💡 Your Question

It seems there are several places in the code where image transforms are defined for Yolo-NAS:

• Predict method on pretrained model:

  yolo_nas = super_gradients.training.models.get("yolo_nas_s", pretrained_weights="coco").cuda()
  model_predictions = yolo_nas.predict(
      "https://deci-pretrained-models.s3.amazonaws.com/sample_images/beatles-abbeyroad.jpg"
  )

  In this case, transforms are defined in default_yolo_nas_coco_processing_params function in processing.py
  There are StandardizeImage and DetectionCenterPadding. No RGB to BGR transform.
• Training with coco_detection_yolo_nas_dataset_params.yaml.
  Validation transforms include StandardizeImage, DetectionPadToSize (which is center padding), and RGB to BGR transform.
• Also, there is a popular notebook for fine-tuning Yolo-NAS on the custom dataset: https://github.com/roboflow/notebooks/blob/main/notebooks/train-yolo-nas-on-custom-dataset.ipynb
  In this notebook coco_detection_yolo_format_train is used for dataset and dataloader creation. Transforms are defined in coco_detection_yolo_format_base_dataset_params.yaml. And there is DetectionPaddedRescale, which includes bottom right padding. No StandardizeImage or RGB to BGR transforms.

So, my question is which transforms should be used for training, validation, and prediction with Yolo-NAS?
Specifically, I'm interested in whether Yolo-NAS accepts RGB or BGR images, whether should images be normalized or not, and what is the proper way of padding (bottom right or center)?
I want to fine-tune the model with the same transforms that were used during training/validation to minimize differences in input data.

Versions

No response