CSPNet: A New Backbone that can Enhance Learning Capability of CNN

CSPNet (Cross Stage Partial Network) is a backbone design aimed at reducing the costly inference computation of convolutional neural networks. The authors trace this cost to duplicate gradient information that arises during network optimization, and they address it by integrating feature maps from the beginning and the end of each network stage, so the network respects the variability of the gradients.

In experiments, CSPNet reduces computation by about 20% while achieving equivalent or even superior accuracy on ImageNet, and it significantly outperforms prior state-of-the-art approaches on AP50 for MS COCO object detection. Because the module is easy to implement and general enough to work with ResNet, ResNeXt, and DenseNet architectures, it offers a way to bring advanced vision models to devices with limited computation resources.