Skip to main content

Best practices

Image capture​

Creating a robust dataset starts with proper image capture techniques.

Match production conditions​

Training images should closely resemble the images your model will encounter in production. This ensures the model learns relevant features and patterns.

Maximize variability​

Capture as much variability as possible that may occur in production scenarios:

  • Different lighting conditions
  • Multiple object positions
  • Edge cases and uncommon scenarios ( but realistic ones)

This diversity ensures your model can handle real-world variations.

Dataset creation​

Maintain image uniqueness​

Avoid Duplicates

Do not add duplicate or very similar images to your dataset. Duplicates can bias your model and waste computational resources.

Separate test sets​

Keep a completely independent subset of images for testing with OneVision:

✅ Use images from different sessions or time periods
❌ Do not use consecutive images from the same sequence
✅ Ensure test data is never seen during training

Strategic validation sampling​

When selecting validation images from a folder:

❌ Don't: Take consecutive images
✅ Do: Use proper sampling techniques (e.g., every nth image, random selection)

This prevents data leakage and ensures honest performance metrics.

Handle questionable images​

Images with unclear or ambiguous labels can harm model performance. Best approach:

  • Avoid adding questionable images from the start
  • If already added, mark them to be ignored during training
  • Review and relabel when possible

Refinement datasets​

Continuous Improvement

Once a model is in production, if you want to refine it, focus your efforts on edge cases:

✅Only add images that caused errors or misclassifications - This targeted approach improves model weaknesses efficiently.

❌Avoid adding random production images that the model already handles well.

Quick checklist​

When creating your dataset, ensure the following best practices are followed:

  • Training images match production conditions
  • Dataset includes diverse scenarios and variations
  • No duplicate or near-duplicate images
  • Test set is completely independent
  • Validation images are properly sampled (not consecutive)
  • Questionable images are excluded or ignored
  • (Refinement datasets) Only error-causing images added

Image Upload Guide - Learn how to upload your images.

Creating a Dataset - Step-by-step instructions to create a dataset and choose the right type balancing.