conv4, 384, /2

unpool, x2

unpool, x2

unpool, x2

concat

concat

concat

conv 1x1,  128

conv 3x3,  128

conv 1x1,  64

conv 1x1,  32 

conv 3x3,  32

conv 3x3,  64

conv 3x3,  32

conv 1x1, 1

conv 1x1, 1

conv 1x1,  4

conv3, 256, /2

conv2, 128, /2

conv4, 64, /2

Image

7x7, 16, /2

\[f_1\]

\[h_1\]

\[f_2\]

\[f_3\]

\[f_4\]

\[h_2\]

\[h_3\]

\[h_4\]

4 distances

1 rotation angle

confidence score

conv1_2

pool_1

conv2_2

pool_2

conv3_3

pool_3

conv4_3

pool_4

conv5_3

pool_5

fc_6

fc_7

upsample

concat

upsample

upsample

concat

concat

concat

Input Image

conv 1x1, 2

conv 1x1, 16

conv 1x1, 2(/16)

conv 1x1, 2(/16)

conv 1x1, 2(/16)

conv 1x1, 2(/16)

conv 1x1, 2(/16)

text/non-text classification

Link prediction