Very Deep Convolutional Networks for Large-Scale Image Recognition

This paper studies how the depth of a convolutional neural network affects its accuracy on large-scale image recognition. Rather than using large convolution filters, the authors build increasingly deep networks entirely from very small 3x3 filters, systematically evaluating how far depth can be pushed while keeping the filter design simple and uniform.

The key finding is that increasing depth to 16-19 weight layers produces a significant accuracy improvement over prior architectures. These deep networks formed the basis of a submission that won first place in localisation and second in classification at the ImageNet Challenge 2014, and the learned representations transferred well to other datasets, achieving state-of-the-art results; the authors released their two best models publicly, which had lasting influence as a standard backbone in computer vision.