notesum.ai

Published at November 25

Utilizing Uncertainty in 2D Pose Detectors for Probabilistic 3D Human Mesh Recovery

cs.CV

Released Date: November 25, 2024

Authors: Tom Wehrbein¹, Marco Rudolph¹, Bodo Rosenhahn¹, Bastian Wandt²

Aff.: ¹Leibniz University Hannover; ²Linkoping University

Arxiv: http://arxiv.org/abs/2411.16289v1

Refer to caption

			3DPW (14)			EMDB (24)
	Models	MPJPE $\downarrow$	PA-MPJPE $\downarrow$	PVE $\downarrow$	MPJPE $\downarrow$	PA-MPJPE $\downarrow$	PVE $\downarrow$
Best of 1	Biggs et al. [4] NeurIPS’20	93.8	59.9	-	-	-	-
	Sengupta et al. [58] CVPR’21	97.1	61.1	-	-	-	-
	ProHMR [38] ICCV’21	97.0	59.8	-	-	-	-
	HierProbHuman [57] ICCV’21	84.9	53.6	-	-	-	-
	HuManiFlow* [59] CVPR’23	83.1	53.9	98.6	113.9	76.4	133.0
	BEDLAM-CLIFF [6] CVPR’23	66.9	43.0	78.5	98.0	60.6	111.6
	ScoreHypo [77] CVPR’24	72.4	44.5	84.6	112.4*	77.9*	131.5*
	ProHMR^† [38]	71.7	43.5	84.7	98.2	60.4	114.5
	Ours	$\bm{62.2}$	$\bm{40.9}$	$\bm{73.9}$	$\bm{84.3}$	$\bm{56.1}$	$\bm{97.7}$
Best of 100	Biggs et al. [4] NeurIPS’20	74.6	48.3	-	-	-	-
	Sengupta et al. [58] CVPR’21	84.4	52.1	-	-	-	-
	ProHMR [38] ICCV’21	81.5	48.2	-	-	-	-
	HierProbHuman [57] ICCV’21	70.9	43.8	-	-	-	-
	HuManiFlow* [59] CVPR’23	64.5	40.0	75.5	88.7	56.5	100.6
	ScoreHypo [77] CVPR’24	63.0	37.6	73.4	87.4*	58.5*	99.6*
	ProHMR^† [38]	55.0	34.7	65.7	76.7	47.1	87.3
	Ours	$\bm{46.2}$	$\bm{29.8}$	$\bm{54.4}$	$\bm{63.6}$	$\bm{40.9}$	$\bm{72.0}$