Created by: fengyuentau

The PyramidBox paper indicates that,

We use the base convolution layers and extra convolutional layers in S^3FD as our backbone layers, which keep layers of VGG16 from conv1_1 to pool5, then convert fc6 and fc7 of VGG16 to conv_fc layers, and then add more convolutional layers to make it deeper.

However, from this line of source code and the definition of conv_block, we can tell that the dilation conv is not applied as S^3FD or DeepLabv1.

Is it a miss or dilation conv not working well?