The effect of depth data and upper limb impairment on lightweight monocular RGB human pose estimation models

Table 1 Dite and MHP prediction errors (MPJPE) on testing data

	Trained with CMU (from scratch)		Trained with H3.6 M (weights loaded from CMU training)		Trained with SUD (weights loaded from H3.6 M training)
Tested on	CMU	H3.6M	CMU	H3.6M	SUD
2D HPE using Dite-HRNet (MPJPE_pix)
RGB	25.70	114.33^*	313.07⁺	3.90	11.86 ± 5.33
CH	18.86	94.55^*	285.92⁺	4.31	11.00 ± 2.80
Cat	22.88	162.70^*	256.26⁺	5.34	15.27 ± 6.47
Fuse	27.73	146.04^*	294.54⁺	18.56	16.48 ± 7.96
3D HPE using MobileHumanPose (MPJPE_mm)
RGB	12.62	463.95^*	130.92⁺	79.67	72.52 ± 28.16
CH	12.38	504.32^*	149.13⁺	72.36	62.79 ± 15.75
Cat	12.27	323.71^*	108.28⁺	114.06	68.00 ± 14.30
Fuse	13.70	415.17^*	131.71⁺	156.39	148.52 ± 18.73

^* indicates results on environments that the model has not encountered before. ⁺ indicates results obtained when testing on data collected in an environment that was previously used for training (prior to fine-tuning)

ISSN: 1475-925X