Motion Recognition Based on Deep Learning and Human Joint Points

Table 4

Several quantization operation modes are supported by TFLite.

Quantitative method	Advantage	Applicable hardware
Weight quantization	The model is reduced by 4 times and accelerated by 2-3 times, with good accuracy	CPU
All integer quantization	The model is reduced by 4 times and accelerated by more than 3 times	CPU, edge TPU, etc.
Float 16 quantization	The model is reduced by 2 times, which has the potential of GPU acceleration	CPU/GPU