DeepFuse neural networks

Xiu Li, Rujiao Long, Kun Jin

Abstract: Generally, pre-trained backbone convolutional models for image classification are directly used as default backbone model for other tasks including detection and segmentation, so as to avoid training a new model from scratch. However, segmentation and detection frameworks need combining low-level and high-level features to boost performance. In this paper, we point out and prove that a simple fusion of low-level and high-level features is insufficient and defined an operation named DeepFuse as a generic family of building blocks for feature fusion. This building block can be plugged into any existing architectures and prevent overfitting effectively for region-level and pixel-level tasks. Based on the Mask R-CNN framework, we achieved a 1.9 AP improvement on the validation set of cityscape and a 1.7 AP improvement on the test set by the use of DeepFuse operation.