Motion Recognition Based on Deep Learning and Human Joint Points

Table 6

Comparison of test accuracy of three models on coco2017 dataset.

Model	PoseLite	Direct conversion model	Weight quantization model
Test data set	2017 val	2017 val	2017 val
AP@0.5 : 0.95	36.9	35.9	27.6
AP@0.5	64.0	62.8	53.8
AP@0.75	35.0	33.4	26.1
AP medium	30.1	28.1	28.2
AP large	46.0	44.0	30.8
AR@0.5 : 0.95	41.5	40.9	32.6
AR@0.5	66.1	64.7	56.9
AR@0.75	40.7	38.7	29.0
AR medium	31.5	29.9	28.9
AR large	55.6	53.4	35.7