OFVL-MS++: Once for visual localization across multiple scenes via a two-stage framework

Learning-based scene coordinate regression (SCoRe) methods have demonstrated significant potential in visual localization. Recently, OFVL-MS has been proposed to resolve the issue of growing model size when the scene number increases. However, the alternating optimization of scene-shared and scene-specific parameters results in inadequate encoding of scene-general and scene-related attributes among scenes. Towards this end, we develop OFVL-MS++, a substantial extension of OFVL-MS, which decouples the localization of multiple scenes into two stages to achieve the maximum efficacy for scene-shared parameters and scene-specific parameters. The first stage seeks to learn the scene-general features by taking the images under various scene domains as inputs and outputting the coarse localization. The second stage aims to learn to adaptively expand the scene-specific parameters for each scene in the deep model, where the requisite positions for expanding scene-specific parameters are automatically learned. Comprehensive experiments show the excellent localization ability of OFVL-MS++ families.