Cropping training images along with their respective annotations

[andualemw1]

Search before asking

• I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

As we know, YOLO supports both square and rectangular images. However, for speed and dataset size considerations, I want to crop an image from 1280x1280 to 640x640. YOLO annotations/labels are originally created based on the image’s width and height. How can I bridge the gap in training the dataset before and after cropping the image while keeping the annotations unchanged?

thanks in advance!

Additional

No response

[UltralyticsAssistant]

👋 Hello @andualemw1, thank you for your interest in YOLOv5 🚀! An Ultralytics engineer will assist you soon.

To get started with cropping images while retaining annotations, you might find our ⭐️ Tutorials helpful. You can explore guides for tasks such as Custom Data Training where managing image sizes and annotations is discussed.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us debug it.

For custom training questions, provide as much information as possible, including dataset image examples and training logs. Verify you are following our Tips for Best Training Results.

Requirements

Ensure you have Python>=3.8.0 with all requirements.txt installed, including PyTorch>=1.8. To get started, run:

git clone https://github.com/ultralytics/yolov5  # clone
cd yolov5
pip install -r requirements.txt  # install

Environments

YOLOv5 can be run in verified environments, including:

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all YOLOv5 GitHub Actions Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training, validation, inference, export, and benchmarks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

Introducing YOLOv8 🚀

Explore our latest object detection model, YOLOv8 🚀! Designed for speed and accuracy, perfect for a wide range of tasks. Discover more in our YOLOv8 Docs and get started with:

pip install ultralytics

Feel free to provide further details to help us address your question! 🔍

[pderrenger]

@andualemw1 to crop your images from 1280x1280 to 640x640 while maintaining correct YOLO annotations, you'll need to adjust the annotations to match the new image dimensions. This involves recalculating the bounding box coordinates based on the new crop position. You can automate this process using a script to update the annotations accordingly. If you need further guidance, consider checking out image processing libraries like OpenCV or PIL for assistance with cropping and recalculating coordinates.

[andualemw1]

Thank you so much for your reply support, this has been my quest for quite a while, I will do accordingly.

[pderrenger]

You're welcome! If you have any more questions as you proceed, feel free to ask.

[andualemw1]

Hello, sorry for coming back again. I am always confused about how to establish YOLOv5 benchmarking. Is it correct to use a confidence threshold of 0.25 on the test dataset for a custom model and for the standard Yolo model?

python val.py --data <data.yaml> --weights <model.pt> --task test --conf-thres 0.25 --iou-thres 0.5

thank you in advance!

[pderrenger]

Yes, your approach to benchmarking with val.py and a confidence threshold of 0.25 is correct. The --conf-thres 0.25 sets the minimum confidence for detections, which is standard for YOLOv5 evaluation. Ensure you're using the same command and thresholds consistently across both custom and standard models for fair comparisons. You can refer to the YOLOv5 documentation for additional details. Let me know if you have further questions!