検出器と検出モデル

Detectors

Detectors find the people in each camera image, so that the feature extractor can then focus on each person's crop.

Detector	Input size	Notes
MediaPipe BlazePose detector	224x224	Extremely lightweight detector. Only use if YOlOX is too slow
YOLOX-tiny	416x416	Great and fast person detector

Feature extractors

The job of feature extractors is much more difficult and important. They are the models that actually specify the body keypoints. Therefore feature extractors affect motion capture in both accuracy and robustness against occlusions.

Kinoca allows you to choose a wide variety of feature extractors that fit your needs. Choosing the best one comes down to your desired mocap FPS, whether you want finger and face tracking, and your available computation resources.

Basic Body Feature extractors (17 Keypoints)

These feature extractors only expose the 17 keypoints as defined by the COCO dataset.

Extractor	Input size	Notes
MoveNet Lightning	192x192	Lightest option

MediaPipe BlazePose Feature extractors (33 Keypoints)

These feature extractors expose the 17 COCO body keypoints, plus the heels, the tips of the feet, the base of the index and little fingers, the thumbs, and a few more face keypoints (33 in total). The extra hand and foot keypoints give a better estimate of the hand and foot rotations than the basic body extractors, but they do not enable full finger tracking.

Extractor	Input size	Notes
MediaPipe BlazePose Lite	256x256	Fastest BlazePose variant
MediaPipe BlazePose Full	256x256	Good balance of speed and accuracy
MediaPipe BlazePose Heavy	256x256	Most accurate BlazePose variant