RelationConv: Exploring Spatial Relationships in Convolution

Xiudong Wang

Abstract: Convolutional neural networks (CNNs) are the dominant methods of computer vision in the last decade. The convolution operator uses the weight sharing mechanism, which is shift invariant and care about the existence of discriminative features, but is insensitive to the global position of these features. To address this issue, we propose to enhance convolutional network by exploring the spatial relationships (named as RelationConv). Specifically, we adds scale factors to filters on different patches of the input features. The scale factors are computed according to the relationships between all patches. With this implementation, RelationConv can acquire both local and global features simultaneously, and build part-whole representation for visual recognition. Extensive experiments including classification on ImageNet and detection as well as segmentation on COCO demonstrate that the proposed RelationConv achieves consistently outperforms competitive baselines. For instance, RelationConv can improve the accuracy of ResNet18 by +1.3%.