Designing Lightweight Feature Descriptor Networks with Depthwise Separable Convolution

抄録

Extracting feature points and their descriptors from images is one of the fundamental techniques in computer vision with many applications such as geometric fitting and camera calibration, and for this task several deep learning models have been proposed. However, existing feature descriptor networks have been developed with the intention of improving the accuracy, and consideration for practical networks that can run on embedded devices has somewhat been deferred. Therefore, the objective of this study is to devise light feature descriptor networks. To this end, we employ lightweight convolution operations that have been developed for image classification networks (e.g. SqueezeNet and MobileNet) for the purpose of replacing the normal convolution operators in the state-of-the-art feature descriptor network, RF-Net. Experimental results show that the model size of the detector can be reduced by up to 80% compared to that of the original size with only a 11% degradation at worst performance in our final lightweight detector model for image matching tasks. Our study indicates that the modern convolution techniques originally proposed for small image classification models can be effectively extended to designing tiny models for the feature descriptor extraction and matching portions in deep local feature learning networks.