In the IC-Net paper, inference should be original resolution and the network predicted output should be original resolution. However in the code inference.py, model predicts 4x downsampled predicted map which is upsampled by tf.image.resize_bilinear ? Why the implementation follows this ?
I notice that even in the original architecture test only branch has a 4x upsample to produce full resolution predicted map. This implementation does the same.
In the IC-Net paper, inference should be original resolution and the network predicted output should be original resolution. However in the code inference.py, model predicts 4x downsampled predicted map which is upsampled by tf.image.resize_bilinear ? Why the implementation follows this ?