Upgrade your network in-place with deformable convolution

Abstract

Improving the performance of the network on is a topic that all deep learning researchers are working together. More new algorithms are proposed for different tasks. But most of these can't avoid spending a lot of time retraining the network model. Deformable convolution is a convolution structure that can extract better features of objects. This paper proposes a new method that can upgrade the standard convolution part of the network to the deformable convolution in-place, inherit the original model parameters, and reduce the time and computational resource cost for retraining. We analyzed the effects of introducing deformable convolution at different depths of the network on speed and performance. And on the detection and semantic segmentation tasks of the PASCAL VOC and COCO, a lot of experiments were carried out on our methods, and have an effective improvement.