MobileNet V2-SSD 检测头数目解析

Backbone网络结构

id	bottom	layer names	top	kernel	pad	stride	chanel output	output num
0	data	input	/	/	/	/	/	(1, 3, 300, 300)
1	data	conv1 + conv1/bn + conv1/scale + relu1	conv1/bn	3	1	2	32	(1, 32, 150, 150)
2_1	conv1/bn	conv2_1/expand + conv2_1/expand/bn + conv2_1/expand/scale + relu2_1/expand	conv2_1/expand/bn	1	0	1	32	(1, 32, 150, 150)
2_1	conv2_1/expand/bn	conv2_1/dwise + conv2_1/dwise/bn + conv2_1/dwise/scale + relu2_1/dwise	conv2_1/dwise/bn	3	1	1	group	(1, 32, 150, 150)
2_1	conv2_1/dwise/bn	conv2_1/linear + conv2_1/linear/bn + conv2_1/linear/scale	conv2_1/linear/bn	1	0	1	16	(1, 16, 150, 150)
2_2	conv2_1/linear/bn	conv2_2/expand + conv2_2/expand/bn + conv2_2/expand/scale + relu2_2/expand	conv2_2/expand/bn	1	0	1	96	(1, 96, 150, 150)
2_2	conv2_2/expand/bn	conv2_2/dwise + conv2_2/dwise/bn + conv2_2/dwise/scale + relu2_2/dwise	conv2_2/dwise/bn	3	1	2	group	(1, 32, 75, 75)
2_2	conv2_2/dwise/bn	conv2_2/linear + conv2_2/linear/bn + conv2_2/linear/scale	conv2_2/linear/bn	1	0	1	24	(1, 24, 75, 75)
3_1	conv2_2/linear/bn	conv3_1/expand + conv3_1/expand/bn + conv3_1/expand/scale + relu3_1/expand	conv3_1/expand/bn	1	0	1	144	(1, 144, 75, 75)
3_1	conv3_1/expand/bn	conv3_1/dwise + conv3_1/dwise/bn + conv3_1/dwise/scale + relu3_1/dwise	conv3_1/dwise/bn	3	1	1	group	(1, 144, 75, 75)
3_1	conv3_1/dwise/bn	conv3_1/linear + conv3_1/linear/bn + conv3_1/linear/scale	conv3_1/linear/bn	1	0	1	24	(1, 24, 75, 75)
block	conv2_2/linear/bn + conv3_1/linear/bn	block_3_1	block_3_1	/	/	/	/	(1, 24, 75, 75)
3_2	conv3_1/linear/bn	conv3_2/expand + conv3_2/expand/bn + conv3_2/expand/scale + relu3_2/expand	conv3_2/expand/bn	1	0	1	144	(1, 144, 75, 75)
3_2	conv3_2/expand/bn	conv3_2/dwise + conv3_2/dwise/bn + conv3_2/dwise/scale + relu3_2/dwise	conv3_2/dwise/bn	3	1	2	group	(1, 144, 38, 38)
3_2	conv3_2/dwise/bn	conv3_2/linear + conv3_2/linear/bn + conv3_2/linear/scale	conv3_2/linear/bn	1	0	1	32	(1, 32, 38, 38)
4_1	conv3_2/linear/bn	conv4_1/expand + conv4_1/expand/bn + conv4_1/expand/scale + relu4_1/expand	conv4_1/expand/bn	1	0	1	192	(1, 192, 38, 38)
4_1	conv4_1/expand/bn	conv4_1/dwise + conv4_1/dwise/bn + conv4_1/dwise/scale + relu4_1/dwise	conv4_1/dwise/bn	3	1	1	group	(1, 192, 38, 38)
4_1	conv4_1/dwise/bn	conv4_1/linear + conv4_1/linear/bn + conv4_1/linear/scale	conv4_1/linear/bn	1	0	1	32	(1, 32, 38, 38)
block	conv3_2/linear/bn + conv4_1/linear/bn	block_4_1	block_4_1	/	/	/	/	(1, 32, 38, 38)
4_2	conv4_1/linear/bn	conv4_2/expand + conv4_2/expand/bn + conv4_2/expand/scale + relu4_2/expand	conv4_2/expand/bn	1	0	1	192	(1, 192, 38, 38)
4_2	conv4_2/expand/bn	conv4_2/dwise + conv4_2/dwise/bn + conv4_2/dwise/scale + relu4_2/dwise	conv4_2/dwise/bn	3	1	1	group	(1, 192, 38, 38)
4_2	conv4_2/dwise/bn	conv4_2/linear + conv4_2/linear/bn + conv4_2/linear/scale	conv4_2/linear/bn	1	0	1	32	(1, 32, 38, 38)
block	block4_1 + conv4_2/linear/bn	block_4_2	block_4_2	/	/	/	/	(1, 32, 38, 38)
4_3	conv4_2/linear/bn	conv4_3/expand + conv4_3/expand/bn + conv4_3/expand/scale + relu4_3/expand	conv4_3/expand/bn	1	0	1	192	(1, 192, 38, 38)
4_3	conv4_3/expand/bn	conv4_3/dwise + conv4_3/dwise/bn + conv4_3/dwise/scale + relu4_3/dwise	conv4_3/dwise/bn	3	1	1	group	(1, 192, 38, 38)
4_3	conv4_3/dwise/bn	conv4_3/linear + conv4_3/linear/bn + conv4_3/linear/scale	conv4_3/linear/bn	1	0	1	64	(1, 64, 38, 38)
4_4	conv4_3/linear/bn	conv4_4/expand + conv4_4/expand/bn + conv4_4/expand/scale + relu4_4/expand	conv4_4/expand/bn	1	0	1	384	(1, 384, 38, 38)
4_4	conv4_4/expand/bn	conv4_4/dwise + conv4_4/dwise/bn + conv4_4/dwise/scale + relu4_4/dwise	conv4_4/dwise/bn	3	1	1	group	(1, 384, 38, 38)
4_4	conv4_4/dwise/bn	conv4_4/linear + conv4_4/linear/bn + conv4_4/linear/scale	conv4_4/linear/bn	1	0	1	64	(1, 64, 38, 38)
block	conv4_3/linear/bn + conv4_4/linear/bn	block_4_4	block_4_4	/	/	/	/	(1, 64, 38, 38)
4_5	conv4_4/linear/bn	conv4_5/expand + conv4_5/expand/bn + conv4_5/expand/scale + relu4_5/expand	conv4_5/expand/bn	1	0	1	384	(1, 384, 38, 38)
4_5	conv4_5/expand/bn	conv4_5/dwise + conv4_5/dwise/bn + conv4_5/dwise/scale + relu4_5/dwise	conv4_5/dwise/bn	3	1	1	group	(1, 384, 38, 38)
4_5	conv4_5/dwise/bn	conv4_5/linear + conv4_5/linear/bn + conv4_5/linear/scale	conv4_5/linear/bn	1	0	1	64	(1, 64, 38, 38)
block	block_4_4 + conv4_5/linear/bn	block_4_5	block_4_5	/	/	/	/	(1, 64, 38, 38)
4_6	conv4_5/linear/bn	conv4_6/expand + conv4_6/expand/bn + conv4_6/expand/scale + relu4_6/expand	conv4_6/expand/bn	1	0	1	384	(1, 384, 38, 38)
4_6	conv4_6/expand/bn	conv4_6/dwise + conv4_6/dwise/bn + conv4_6/dwise/scale + relu4_6/dwise	conv4_6/dwise/bn	3	1	1	group	(1, 384, 38, 38)
4_6	conv4_6/dwise/bn	conv4_6/linear + conv4_6/linear/bn + conv4_6/linear/scale	conv4_6/linear/bn	1	0	1	64	(1, 64, 38, 38)
block	block_4_5 + conv4_6/linear/bn	block_4_6	block_4_6	/	/	/	/	(1, 64, 38, 38)
4_7	conv4_6/linear/bn	conv4_7/expand + conv4_7/expand/bn + conv4_7/expand/scale + relu4_7/expand	conv4_7/expand/bn	1	0	1	384	(1, 384, 38, 38)
4_7	conv4_7/expand/bn	conv4_7/dwise + conv4_7/dwise/bn + conv4_7/dwise/scale + relu4_7/dwise	conv4_7/dwise/bn	3	1	2	group	(1, 384, 19, 19)
4_7	conv4_7/dwise/bn	conv4_7/linear + conv4_7/linear/bn + conv4_7/linear/scale	conv4_7/linear/bn	1	0	1	96	(1, 96, 19, 19)
5_1	conv4_7/linear/bn	conv5_1/expand + conv5_1/expand/bn + conv5_1/expand/scale + relu5_1/expand	conv5_1/expand/bn	1	0	1	576	(1, 576, 19, 19 )
5_1	conv5_1/expand/bn	conv5_1/dwise + conv5_1/dwise/bn + conv5_1/dwise/scale + relu5_1/dwise	conv5_1/dwise/bn	3	1	1	group	(1, 576, 19, 19)
5_1	conv5_1/dwise/bn	conv5_1/linear + conv5_1/linear/bn + conv5_1/linear/scale	conv5_1/linear/bn	1	0	1	96	(1, 96, 19, 19)
block	block_4_7 + conv5_1/linear/bn	block_5_1	block_5_1	/	/	/	/	(1, 96, 19, 19)
5_2	conv5_1/linear/bn	conv5_2/expand + conv5_2/expand/bn + conv5_2/expand/scale + relu5_2/expand	conv5_2/expand/bn	1	0	1	576	(1, 576, 19, 19 )
5_2	conv5_2/expand/bn	conv5_2/dwise + conv5_2/dwise/bn + conv5_2/dwise/scale + relu5_2/dwise	conv5_2/dwise/bn	3	1	1	group	(1, 576, 19, 19)
5_2	conv5_2/dwise/bn	conv5_2/linear + conv5_2/linear/bn + conv5_2/linear/scale	conv5_2/linear/bn	1	0	1	96	(1, 96, 19, 19)
block	block_5_1 + conv5_2/linear/bn	block_5_2	block_5_2	/	/	/	/	(1, 96, 19, 19)
5_3	conv5_2/linear/bn	conv5_3/expand + conv5_3/expand/bn + conv5_3/expand/scale + relu5_3/expand	conv5_3/expand/bn	1	0	1	576	(1, 576, 19, 19)
5_3	conv5_3/expand/bn	conv5_3/dwise + conv5_3/dwise/bn + conv5_3/dwise/scale + relu5_3/dwise	conv5_3/dwise/bn	3	1	2	group	(1, 576, 10, 10)
5_3	conv5_3/dwise/bn	conv5_3/linear + conv5_3/linear/bn + conv5_3/linear/scale	conv5_3/linear/bn	1	0	1	160	(1, 160, 10, 10)
6_1	conv5_3/linear/bn	conv6_1/expand + conv6_1/expand/bn + conv6_1/expand/scale + relu6_1/expand	conv6_1/expand/bn	1	0	1	960	(1, 960, 10, 10 )
6_1	conv6_1/expand/bn	conv6_1/dwise + conv6_1/dwise/bn + conv6_1/dwise/scale + relu6_1/dwise	conv6_1/dwise/bn	3	1	1	group	(1, 960, 10, 10)
6_1	conv6_1/dwise/bn	conv6_1/linear + conv6_1/linear/bn + conv6_1/linear/scale	conv6_1/linear/bn	1	0	1	160	(1, 160, 10, 10)
block	conv5_3/linear/bn + conv6_1/linear/bn	block_6_1	block_6_1	/	/	/	/	(1, 160, 10, 10)
6_2	conv6_1/linear/bn	conv6_2/expand + conv6_2/expand/bn + conv6_2/expand/scale + relu6_2/expand	conv6_2/expand/bn	1	0	1	960	(1, 960, 10, 10 )
6_2	conv6_2/expand/bn	conv6_2/dwise + conv6_2/dwise/bn + conv6_2/dwise/scale + relu6_2/dwise	conv6_2/dwise/bn	3	1	1	group	(1, 960, 10, 10)
6_2	conv6_2/dwise/bn	conv6_2/linear + conv6_2/linear/bn + conv6_2/linear/scale	conv6_2/linear/bn	1	0	1	160	(1, 160, 10, 10)
block	block_6_1 + conv6_2/linear/bn	block_6_2	block_6_2	/	/	/	/	(1, 160, 10, 10)
6_3	conv6_2/linear/bn	conv6_3/expand + conv6_3/expand/bn + conv6_3/expand/scale + relu6_3/expand	conv6_3/expand/bn	1	0	1	960	(1, 960, 10, 10 )
6_3	conv6_3/expand/bn	conv6_3/dwise + conv6_3/dwise/bn + conv6_3/dwise/scale + relu6_3/dwise	conv6_3/dwise/bn	3	1	1	group	(1, 960, 10, 10)
6_3	conv6_3/dwise/bn	conv6_3/linear + conv6_3/linear/bn + conv6_3/linear/scale	conv6_3/linear/bn	1	0	1	320	(1, 320, 10, 10)
6_4	conv6_3/linear/bn	conv6_4 + conv6_4/bn + conv6_4/scale + relu6_4	conv6_4/bn	1	0	1	1280	(1, 1280, 10, 10)
14_1	conv6_4/bn	conv14_1 + conv14_1/bn + conv14_1/scale + conv14_1/relu	conv14_1	1	0	1	256	(1, 256, 10, 10)
14_2	conv14_1	conv14_2 + conv14_2/bn + conv14_2/scale + conv14_2/relu	conv14_2	3	1	2	512	(1, 512, 5, 5)
15_1	conv14_2	conv15_1 + conv15_1/bn + conv15_1/scale + conv15_1/relu	conv15_1	1	0	1	128	(1, 128, 5, 5)
15_2	conv15_1	conv15_2 + conv15_2/bn + conv15_2/scale + conv15_2/relu	conv15_2	3	1	2	256	(1, 256, 3, 3)
16_1	conv15_2	conv16_1 + conv16_1/bn + conv16_1/scale + conv16_1/relu	conv16_1	1	0	1	128	(1, 128, 3, 3)
16_2	conv16_1	conv16_2 + conv16_2/bn + conv16_2/scale + conv16_2/relu	conv16_2	3	1	2	256	(1, 256, 2, 2)

检测头结构

id	bottom	input num	layer names	top	kernel	pad	stride	chanel output	output num
loc_1	conv4_7/expand/bn	(1, 384, 38, 38)	conv5_mbox_loc	conv5_mbox_loc	1	0	1	16	(1, 4*4, 38, 38)
conf_1	conv4_7/expand/bn	(1, 384, 38, 38)	conv5_mbox_conf	conv5_mbox_conf	1	0	1	cls4x	(1, 4*cls, 38, 38)
box_1	conv4_7/expand/bn	(1, 384, 38, 38)	conv5_mbox_priorbox	conv5_mbox_priorbox	/	/	/	/	size:16~45， 4个
loc_2	conv5_3/expand/bn	(1, 576, 19, 19)	conv11_mbox_loc	conv11_mbox_loc	1	0	1	24	(1, 4*6, 19, 19)
conf_2	conv5_3/expand/bn	(1, 576, 19, 19)	conv11_mbox_conf	conv11_mbox_conf	1	0	1	6*cls	(1, 6*cls, 19, 19)
box_2	conv5_3/expand/bn	(1, 576, 19, 19)	conv11_mbox_priorbox	conv11_mbox_priorbox	/	/	/	/	size:45~99， 6个
loc_3	conv6_4/expand/bn	(1, 1280, 10, 10)	conv13_mbox_loc	conv13_mbox_loc	1	0	1	24	(1, 4*6, 10, 10)
conf_3	conv6_4/expand/bn	(1, 1280, 10, 10)	conv13_mbox_conf	conv13_mbox_conf	1	0	1	6*cls	(1, 6*cls, 10, 10)
box_3	conv6_4/expand/bn	(1, 1280, 10, 10)	conv13_mbox_priorbox	conv13_mbox_priorbox	/	/	/	/	size:99~153， 6个
loc_4	conv14_2	(1, 512, 5, 5)	conv14_2_mbox_loc	conv14_2_mbox_loc	1	0	1	24	(1, 4*6, 5, 5)
conf_4	conv14_2	(1, 512, 5, 5)	conv14_2_mbox_conf	conv14_2_mbox_conf	1	0	1	6*cls	(1, 6*cls, 5, 5)
box_4	conv14_2	(1, 512, 5, 5)	conv14_2_mbox_priorbox	conv14_2_mbox_priorbox	/	/	/	/	size:153~207， 6个
loc_5	conv15_2	(1, 256, 3, 3)	conv15_2_mbox_loc	conv15_2_mbox_loc	1	0	1	24	(1, 4*6, 3, 3)
conf_5	conv15_2	(1, 256, 3, 3)	conv15_2_mbox_conf	conv15_2_mbox_conf	1	0	1	6*cls	(1, 6*cls, 3, 3)
box_5	conv15_2	(1, 256, 3, 3)	conv15_2_mbox_priorbox	conv15_2_mbox_priorbox	/	/	/	/	size:207~261， 6个
loc_6	conv16_2	(1, 256, 2, 2)	conv16_2_mbox_loc	conv16_2_mbox_loc	1	0	1	24	(1, 4*6, 2, 2)
conf_6	conv16_2	(1, 256, 2, 2)	conv16_2_mbox_conf	conv16_2_mbox_conf	1	0	1	6*cls	(1, 6*cls, 2, 2)
box_6	conv16_2	(1, 256, 2, 2)	conv16_2_mbox_priorbox	conv16_2_mbox_priorbox	/	/	/	/	size:261~315， 6个

Prior Box

id	layer name	min size	max size	ratio	box num	size list
1	conv5_mbox_priorbox	16	45	2.0	4	(16,16),(27,27),(23,11),(11,23)
2	conv11_mbox_priorbox	45	99	2.0,3.0	6	(45,45),(67,67),(64,32),(32,64),(78)
3	conv13_mbox_priorbox	99	153	2.0,3.0	6	140,
4	conv16_2_mbox_priorbox	153	207	2.0,3.0	6	216
5	conv15_2_mbox_priorbox	207	261	2.0,3.0	6	293
6	conv16_2_mbox_priorbox	261	315	2.0,3.0	6	369

解析 mbox_loc:

从6层feature map出来的prior box的相对位置结果(x, y, w, h)的拼接

id	layer name	size	output num	box num
1	conv5_mbox_loc_flat	38*38	4 / 16	38x38x4=5776
2	conv11_mbox_loc_flat	19*19	6 / 24	19x19x6=2166
3	conv13_mbox_loc_flat	10*10	6 / 24	10x10x6=600
4	conv14_2_mbox_loc_flat	5*5	6 / 24	5x5x6=150
5	conv15_2_mbox_loc_flat	3*3	6 / 24	3x3x6=54
6	conv16_2_mbox_loc_flat	2*2	6 / 24	2x2x6=24

5776+2166+600+150+54+24=8770
8770*4=35080

mbox_conf_flatten:

id	layer name	size	output num	box num
1	conv5_mbox_conf_flat	38*38	4 / 4*cls_num	38x38x4=5776
2	conv11_mbox_conf_flat	19*19	6 / 6*cls_num	19x19x6=2166
3	conv13_mbox_conf_flat	10*10	6 / 6*cls_num	10x10x6=600
4	conv14_2_mbox_conf_flat	5*5	6 / 6*cls_num	5x5x6=150
5	conv15_2_mbox_conf_flat	3*3	6 / 6*cls_num	3x3x6=54
6	conv16_2_mbox_conf_flat	2*2	6 / 6*cls_num	2x2x6=24

5776+2166+600+150+54+24=8770
8770*cls_num= 8770 cls_num

mbox_priorbox

id	layer name	size	output num	box num
1	conv5_mbox_priorbox	38*38	4 / 16	38x38x4=5776
2	conv11_mbox_priorbox	19*19	6 / 24	19x19x6=2166
3	conv13_mbox_priorbox	10*10	6 / 24	10x10x6=600
4	conv14_2_mbox_priorbox	5*5	6 / 24	5x5x6=150
5	conv15_2_mbox_priorbox	3*3	6 / 24	3x3x6=54
6	conv16_2_mbox_priorbox	2*2	6 / 24	2x2x6=24

5776+2166+600+150+54+24=8770
8770*4=35080
(1, 2 35080)