Very Deep Convolutional Networks for Large-Scale Image Recognition (VGG)

Demonstrated that depth using homogeneous stacks of tiny 3x3 convolution filters is critical for high visual representation capacity.

The standard feature extractor for perceptual loss functions in neural style transfer, super-resolution, and image synthesis.

Assumptions

• Two stacked 3x3 conv layers have an effective receptive field of 5x5 with fewer parameters and more non-linearities

Limitations

• High parameter count (138M in VGG16) and massive fully connected heads leading to slow inference and high memory bandwidth