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

Table 5 Model training parameters

	Dite-HRNet	MobileHumanPose
Epochs	10*	25
Learning rate	1e-3	1e-3 (decreased by a factor of 10 after the 17th and 25th epoch)
Batch size	16	64
Input shape	384 × 288	256 × 256
Tensor normalization	Mean: (0.485, 0.456, 0.406, 0.500) Standard deviation: (0.229, 0.224, 0.225, 0.366)	Mean: (0.485, 0.456, 0.406, 0.500) Standard deviation: (0.229, 0.224, 0.225, 0.366)
Number of workers	2	40
Number of GPUs	4	4

^*The original Dite-HRNet model was trained with 270 epochs, which was not feasible in this study. Pilot training revealed that, when training Dite-HRNet with the full CMU data set, the model was done most of its learning within the first epoch. Therefore, a standard 10 epochs of training was set for this model

ISSN: 1475-925X