Modern instance segmentation¶

Outputs each object's class, box and instance mask. Datasets require contour polygon annotations; ordinary horizontal boxes cannot replace instance contours.

Defaults: input 640, maximum epochs 1000, batch size 16, workers 8, optimizer auto. See modern common parameters for basic settings, geometric augmentation and loss weights.

Segmentation copy_paste uses native mask copy-paste. Background erasing avoids objects using their bounding boxes. The UI exposes no separate mask-loss coefficient, mask downsampling or multi-GPU list; do not confuse engine-level advanced parameters with available controls.

Charts distinguish box and mask validation metrics: (B) denotes boxes and (M) masks. Values come from actual validation.

Training results are stored in runs inside the dataset directory; ONNX and class deployment configuration are exported automatically. Select instance segmentation weights for inference to obtain the corresponding masks.

See the segmentation parameter table for all control defaults and ranges.