Scale-invariant Convolutional Capsule Network

Abstract

Scale-invariant feature detection plays an important role in computer vision. Inspired by capsule networks and Hough transform, we presented a scale-invariant feature detection and extraction module that is streamlined, lightweight, and interpretable. We reduced the overhead by combining the learnable feature detector with an efficient scale aware parameter voting mechanism. Also, a routing mechanism can be added to further refine the extracted features to boost performance. Compared with popular multicolumn or feature pyramid based methods, our proposed method is more lightweight both parameter wise and architectural wise, while maintaining good multiscale performance, especially in intensively scale transformed scenarios. Meanwhile, its streamlined sequential pipeline makes it easy to integrate into other models in a plug-and-play manner.