. Author manuscript; available in PMC: 2017 Feb 1.

Published in final edited form as: IEEE Trans Pattern Anal Mach Intell. 2015 Sep 23;38(8):1640–1650. doi: 10.1109/TPAMI.2015.2481404

TABLE 7.

Results on the UCF human action database. Comparative results with state-of-the-art methods are obtained using a subset of 10 joints, which can be computed extremely fast. The derived approach outperforms state-of-the-art algorithms when using the full set 15 joints. CRF: Conditional Random Fields. TS: Temporal Segment. JAS: Joint Angle Similarity. RW: Random Walk.

Method	Accuracy	Time/Sample
CRF [15]	94.29%	–
BoW [15]	94.06%	–
TS [15]	95.94%	–
JAS [38]	97.37%	–
RW + SVM [24]	97.16% ± 0.26%	2.18 s
Path Kernel + RKLR (l₁)	97.99% ± 0.16%	–
Path Kernel + RKLR (l₂)	97.78% ± 0.22%	–
Path Kernel + SVM	98.70% ± 0.16%	1.76 s