Leaning Local Features for Visual Localization Using Images Generated by Diffusion-Based Model

Long-term Visual Localization is a challenging task, and robustness of local features to changes in imaging conditions is important. A common approach to learning local features is to warp a single image using various homographic transformations to obtain many images from different viewpoints to increase the training data. However, this method cannot simulate changes in imaging conditions such as time, season, and weather. In this paper, we improve the robustness of local features by using an image generation model to transform images to different imaging conditions and simulate changes in viewpoints and imaing conditions simultaneously. Using state-of-the-art image generation models, we generate images of various imaging conditions from a single real image by faithfully reproducing complex changes in appearance in the real world. We trained deep learning feature point extraction models using datasets augmented by these methods. Results evaluated on existing Visual Localization benchmarks show improved robustness in nighttime pose estimation.