Modern oriented object detection¶

For elongated objects with a clear orientation, tilted parts and similar tasks, producing oriented boxes. Select OBB-specific weights; each data row must contain a class ID and four normalized corners, for nine columns total.

The software checks corner ranges, class IDs, ordering, geometric validity and the presence of valid training objects. Ordinary five-column box labels and unconverted raw DOTA text are not accepted.

Defaults: input 640, maximum epochs 1000, batch size 16, workers 0 to reduce Windows data loading compatibility issues. Enable parallel loading according to your actual environment when needed.

Do not confuse the two angle controls:

Parameter	Meaning
degrees	Maximum absolute rotation angle for training image augmentation
angle	Model angle loss weight, default 1; not a rotation angle

This page does not show standard detection copy-paste or background erasing controls. Other basic parameters and geometric augmentation follow modern detection, limited to losses actually supported by the current model architecture.

Automatic ONNX export uses nms=False and simplify=False to prevent native OBB embedded-NMS export from re-enabling simplification. External deployment requires OBB-specific oriented-box decoding and postprocessing; do not treat it as ordinary horizontal-box output. Deployment configuration includes task: obb and angle_unit: radian.

See the OBB parameter table for complete ranges.