Fig. 3.

FPN with Inception-v4 encoder used for regression. In the encoding branch, the image is processed with the Inception-v4 classification network. Skip connections (dashed lines) are inserted after layers where the output was reduced in spatial size by factors of 4, 8, 16, and 32, respectively. The skip connections feed FPN blocks where they undergo a series of convolutions. The outputs are upsampled independently by factors of 1, 2, 4, and 8, respectively. Their outputs are added and inserted in a spatial dropout layer activated only during training for regularization purposes. After dropout, convolutions are followed by an upsampling by a factor of 4 to match the original image size and the sigmoid activation function