Yolo 11 image preprocess

Support YOLO
, ,
1

Hello,

After extensive reading and investigation, there are still some parts of the preprocessing pipeline in YOLO11 that I’m struggling to fully understand. I would greatly appreciate your clarification.

Specifically, before an image is passed into the model, does it undergo letterbox resizing or center cropping? How does this behavior differ between training, validation, and testing phases? In addition, how are the arguments rect and imgsz related?
I would also like to know whether the process is different between classification and detection models.

Thank you in advance for your help in clarifying these points.

2

It will help a lot to understand your goal. Your questions are clear enough to answer, but understanding your true goal will help to give answers to your questions that are truly meaningful.

For your questions:

  1. Preprocessing for the classify and detect tasks or models are slightly different.
  2. The most relevant relation to imgsz and rect mostly boils down to what the largest dimension is for a non-square image, aside from that, they have very different functions.
  3. Preprocessing of images with letterbox resizing or center cropping depends on several factors, hence understanding your goals will help a great deal in providing you a sensible answer.
    • There might be slight differences for validation or inference, depending on the model task, the model format (file type), and the arguments used.

Following the code will help answer all of these questions. Starting with the BasePredictor in ultralytics/engine/predictor.py you will find a .preprocess() method. This is used for all the tasks for .predict() calls, unless the method is overwritten. For the detect models, you can see there is no override to the .preprocess() method in ultralytics/models/yolo/detect/predict.py and since the DetectionPredictor inherits from BasePredictor, it will use the BasePredictor.preprocess() method. When calling .predict() for a classify task, you’ll find that in ultralytics/models/yolo/classify/predict.py the ClassificationPredictor has an override for the .preprocess() method, so the code executed is different for the classify models. This pattern is repeated for the various models/tasks in the code, so you should be able to find everything you want to know about preprocessing directly in the source code. You can find the class each model or task uses in the task_map of the YOLO class.