Non-local pose means for denoising motion capture data

Motion capture is commonly used in movies and games, and may become more widespread if anticipated virtual reality and augmented reality applications become popular. Unfortunately, body motion capture generally suffers from significant noise resulting from intrinsic factors not present in other media such as images and videos. This paper adapts the nonlocal means (NLM) principle from image processing to noise reduction in motion capture data. We show that the NLM principle can be applied while respecting the rotational nature of skeletal movement by operating in the vector space obtained by a log/exponential map. NLM has been previously overlooked in the field of motion data. The results are compellingly effective motion denoising for small and large amounts of Gaussian noise. Our results rival or outperform other techniques in a survey of standard denoising methods in signal processing.