ResNet: Deep Residual Learning for Image Recognition
Winner in several tracks of the ILSVRC & COCO 2015 competitions: ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation. The main novelty of this paper is the use of residual blocks which is a series of 2 or 3 conv layers with a skip connection.
The skip connection works as an identity shortcut to skip a block when it is not useful for a particular class. This allows to build much deeper architectures.
The authors introduced 5 architectures with respectively 18, 34, 50, 101 and 152 layers.