Handling small objects during training

[KDeser]

Sometimes objects can be collapsed down to almost nothing when resizing large images for training. This is problematic and leads to poor performance because the model attempts to learn from these tiny smudges of pixels.

I suggest implementing some logic within the data loader that removes labels where the bboxes are below a certain size, or crops the image so that the objects remain larger.

[UltralyticsAssistant]

👋 Hello @KDeser, thank you for sharing your thoughts and contributing to the Ultralytics project 🚀!

We understand the challenges that arise when dealing with small objects during training and appreciate your suggestions for handling this issue. Resizing and bounding box handling are crucial for improving model performance.

If this is a ❓ Question or Feature Request, please feel free to elaborate further on your use case or share any specific examples/insights so we can better understand your needs.
If this is a 🐛 Bug Report exposing an issue in the current implementation, we kindly ask you to provide a minimum reproducible example along with additional information like your dataset, parameters, and environment setup. This will help us debug the issue more effectively.

To explore potential solutions, make sure to check out our Docs for training tips and preprocessing techniques. For custom training specifics, verify that you're following our Tips for Best Training Results, which might include strategies related to small object detection.

Upgrade

Before proceeding further, ensure you’ve upgraded to the latest version of the ultralytics package and dependencies to rule out the possibility of an already-resolved issue:

pip install -U ultralytics

Community Resources

We also encourage you to connect with the Ultralytics community to discuss ideas and gather feedback:

• Live chat on Discord 🎧
• In-depth discussions on Discourse
• Knowledge sharing on our Subreddit

Training Environments

YOLO can be run on various verified environments, including:

• Notebooks with GPUs:
• GCP Deep Learning VMs (Quickstart)
• AWS Deep Learning AMIs (Quickstart)
• Docker Images (Quickstart)

Status

When this badge is green, all Ultralytics CI tests are passing, verifying the correct operation of YOLO across all platforms.

This is an automated response, but an Ultralytics engineer will review your discussion and provide further assistance soon. Thank you for your patience 😊!

[glenn-jocher]

@KDeser to address small object challenges during training, you can increase the imgsz parameter or try P2 model variants.

[robmarkcole]

Can min_wh be used in eval/test too? I am interested in metrics for very small objects vs average size