Model Lightweighting for Real-time Distraction Detection on Resource-Limited Devices

Table 1

The architecture of MobileNetV2-tiny transforming the bottleneck residual block from to , with channels ranging from to .

Layer	Operator	Input
Conv1	Conv2d	224 × 224 ×   3	1	1	32	22
Conv2	Bottleneck	112 × 112 ×   32	1	1	16	11
Conv3	Bottleneck	112 × 112 ×   16	2	2	24	17
Conv4	Bottleneck	56 × 56 ×   24	3	2	32	22
Conv5	Bottleneck	28 × 28 ×   32	4	1	64	45
Conv6	Bottleneck	14 × 14 ×   64	3	3	96	67
Conv7	Bottleneck	14 × 14 ×   96	3	4	160	144
Conv8	Bottleneck	7 × 7 ×   160	1	2	320	288
Conv9	Conv2d	7 × 7 ×   320	1	1	1280	1152
AvgPool	AvgPool	7 × 7 ×   1280	1	1	—	—
Fc	Linear	1280  10	—	—	—	—