US 12,327,338 B2
Methods for improving datasets for skeleton-based action detection
Marios Savvides, Pittsburgh, PA (US); Yu Kai Huang, Pittsburgh, PA (US); and Eddie Yu, Pittsburgh, PA (US)
Assigned to Carnegie Mellon University, Pittsburgh, PA (US)
Filed by CARNEGIE MELLON UNIVERSITY, Pittsburgh, PA (US); and AnyVision Interactive Technologies, Ltd., Holon (IL)
Filed on Oct. 20, 2023, as Appl. No. 18/491,258.
Claims priority of provisional application 63/419,118, filed on Oct. 25, 2022.
Claims priority of provisional application 63/417,820, filed on Oct. 20, 2022.
Prior Publication US 2024/0135503 A1, Apr. 25, 2024
Prior Publication US 2024/0233095 A9, Jul. 11, 2024
Int. Cl. G06T 5/70 (2024.01); G06T 3/4007 (2024.01); G06T 3/60 (2024.01); G06T 5/20 (2006.01); G06T 7/20 (2017.01); G06T 7/60 (2017.01)
CPC G06T 5/70 (2024.01) [G06T 3/4007 (2013.01); G06T 3/60 (2013.01); G06T 5/20 (2013.01); G06T 7/20 (2013.01); G06T 7/60 (2013.01); G06T 2207/20032 (2013.01); G06T 2207/20044 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A method of denoising an action dataset comprising video clips of actions comprising:
breaking the video clips into one or more fixed length clips of a predetermined number of frames;
determining the number of frames in the fixed number of frames that are noisy;
determining a ratio of the number of noisy frames to the total number of frames in the fixed length clip; and
removing the fixed length clip from the dataset when the ratio exceeds a predetermined noise threshold;
wherein skeletal representations of one or more persons in the fixed length clips are extracted from each frame of the fixed length clips, the skeletal representations comprising a set of joint coordinates defining locations of joints of the skeletal representation in a coordinate system defined in the context of each frame.