Rethinking Low-level Features for Interest Point Detection and Description

Although great efforts have been made for interest point detection and description, the current learning-based methods that use high-level features from the higher layers of Convolutional Neural Networks (CNN) do not completely outperform the conventional methods. On the one hand, interest points are semantically ill-defined and high-level features that emphasize semantic information are not adequate to describe interest points; On the other hand, the existing methods using low-level information usually perform detection on multi-level feature maps, which is time consuming for real time applications. To address these problems, we propose a Low-level descriptor-Aware Network (LANet) for interest point detection and description in self-supervised learning. Specifically, the proposed LANet exploits the low-level features for interest point description while using high-level features for interest point detection. Experimental results demonstrate that LANet achieves state-of-the-art performance on the homography estimation benchmark. Notably, the proposed LANet is a front-end feature learning framework that can be deployed in downstream tasks that require interest points with high-quality descriptors.