Modern object detection

For tasks producing ordinary horizontal rectangular boxes, object classes and confidence scores. Select preset detection weights and an ordinary box-label dataset.

  1. Select model weights and the dataset root.
  2. Set single input size, batch size, epochs, device and other common parameters.
  3. Adjust rotation, flipping, translation, scaling and Mosaic for the task. Copy-paste and background erasing use annotation-safe adapters; see the common parameters page for their exact behavior.
  4. View training loss and actual validation metrics such as Precision, Recall and mAP. Only values returned by the engine are displayed; accuracy is not inferred from training loss.
  5. After completion, inspect checkpoints, ONNX and deployment class configuration in the experiment's weights; see inference and export.

Defaults: input size 640, maximum epochs 1000, batch size 16, workers 8, patience 100, optimizer auto. Exact ranges and augmentation defaults are in the parameter checklist.