Learning to Predict Stability of Visual Features

Visual SLAM usually stores many features and descriptors in the mapping phase. In the case of long-term visual SLAM, however, a large amount of the features in the map are unstable. In this paper, we propose a new method to predict the cross-seasonal stability of visual features. Our method is easy to incorporate with the mapping phase and retains those visual features with high repeatability, an essential property in the face of irregular environmental changes. We train the network and conduct experiments on a long-term outdoor data set. The experimental results show that our method can maintain a good prediction in different environments.