Inference for Keypoints

Pose and landmark models label each sample with a Keypoints field, which you generate by passing a model to apply_model().

The Ultralytics integration shows this in practice: a YOLO pose model can be passed directly to apply_model() with no wrapping required, producing keypoints that render with FiftyOne’s built-in skeleton support.

See Ultralytics keypoint inference

For a fully custom model, configure a TorchImageModel with a KeypointDetectorOutputProcessor so it works with apply_model() directly, as described in Inference with custom models.