v1.4.7

What's new in 1.4.7

Fixed

• GPU detection — _utils.py never imported cv2, so CUDA was always reported absent. GPU acceleration now works on a CUDA-enabled OpenCV: pose(device='gpu') (OpenPose), flow.dense(use_gpu=True), flow.sparse(use_gpu=True).
• GPU sparse optical flow path corrected (point shapes + calc() returns).
• MgList.show(key='mgx'/'vgx'/…) on motiongrams/videograms no longer crashes — selects the right panel.
• Audio methods inherited by MgVideo (e.g. video.spectrogram()) no longer fail on a missing cache attribute.
• CI on macOS/Windows: skip .ogg tests without libtheora; force the Matplotlib Agg backend.

Added

• pose(use_cache=True) — re-render a different style/overlay/background from cached keypoints without re-running inference.
• pose(data_format='c3d') — export markers to a C3D motion-capture file (optional c3d dependency).
• tempogram(onset_strength=False) — single-panel tempogram matching the chromagram size.
• beat_statistics(source='motion') on MgVideo — timing statistics from movement onsets.

Changed

• Average-pose labels: numbers only (normalised QoM 0–1 | dominant frequency), stronger de-overlap.
• motiontempo(): QoM normalised 0–1, renamed "Motion spectrum", with average QoM and beat-frequency lines + numbers.