Very Deep Convolutional Networks for Large-Scale Image Recognition .

Abstract

VGGNet demonstrated that depth matters. By using small 3x3 filters but stacking them deep (16-19 layers), they achieved state-of-the-art performance on ImageNet.

Key Insight

Stack of three 3x3 conv layers has same receptive field as one 7x7 layer, but:

• More non-linearities (deeper)
• Fewer parameters

Legacy

VGG-16 and VGG-19 became the go-to feature extractors for transfer learning before ResNet.