Skip to content

AccuracyAndConfiguration.md

Latest commit

 

History

History
61 lines (45 loc) · 3.98 KB

AccuracyAndConfiguration.md

File metadata and controls

61 lines (45 loc) · 3.98 KB

💡 Accuracy Depends on Configuration

Detection accuracy is not automatic — it is the result of matching inference-time configuration to how the model was trained.
Even a correctly trained model can produce poor results if preprocessing, resize behavior, or thresholds do not align with training assumptions.

There is no one-size-fits-all configuration. Optimal settings depend on your dataset, training pipeline, and application requirements.

🔑 Key Factors That Affect Accuracy

  1. Image Preprocessing & Resize Mode

    • Controlled via ImageResize.
    • Resize behavior must match the preprocessing used during training. A mismatch here is one of the most common causes of degraded detection accuracy.
      Proportional dataset Stretched dataset
      proportional stretched
      Use ImageResize.Proportional (default) when the training dataset preserved aspect ratio (letterboxing or padding). Use ImageResize.Stretched if images were resized directly to the model input size (e.g. 640×640) without preserving aspect ratio.

    ℹ️ Most Ultralytics YOLO models are trained using proportional resizing with padding (letterboxing). ImageResize.Proportional is the default in YoloDotNet and is the correct choice in the majority of cases.

    Important:
    Selecting the wrong resize mode can reduce detection accuracy. Change this only if you know your training pipeline resized images without preserving aspect ratio.

  2. Sampling Options

    • Controlled via SamplingOptions.

    • Define how pixel data is resampled when resizing (e.g., Cubic, NearestNeighbor, Bilinear). This choice has a direct impact on the accuracy of your detections, as different resampling methods can slightly alter object shapes and edges.

    • YoloDotNet default:

      SamplingOptions = new SKSamplingOptions(SKFilterMode.Nearest, SKMipmapMode.None);

      This default prioritizes deterministic behavior and performance while matching common training pipelines. Changing the sampling method may improve visual quality, but can subtly affect detection accuracy.

      💡 Tip: Check the ResizeImage Benchmarks for examples of different SamplingOptions and to help you choose the best settings for your needs.

  3. Confidence & IoU Thresholds

    Detection results are filtered using thresholds you configure at inference time:

    • Confidence
      Minimum probability required for a detection to be considered valid.
    • IoU (Intersection-over-Union)
      Controls how overlapping detections are merged or suppressed.

    General guidance:

    • Too low → more false positives.
    • Too high → more missed detections.

    Fine-tune these values based on your dataset, object density, and tolerance for false positives.

💡 Recommended approach: Start with the defaults, verify results on a representative validation set, then adjust ImageResize, SamplingOptions, and Confidence/IoU thresholds incrementally.

Visualization & Styling (Optional)

Visualization does not affect inference accuracy, but YoloDotNet provides full control over how results are rendered.

  • Custom Colors – use built-in class colors or define your own
  • Font Style & Size – fully configurable label rendering
  • Custom Fonts – load your own font files for overlays

For advanced customization, see the extension methods in the main YoloDotNet repository — they provide a solid, production-ready foundation.

Practical drawing and rendering examples are available in the demo project source code.