DiffSegLoc: Privacy Preserving Visual Localization with Diffusion Models and Segmentation Consistency

Visual localization is fundamental in computer vision and is widely integrated in autonomous driving, virtual reality and industrial manufacturing. Robust and accurate visual localization has attracted intense attention. In practice, query images often contain privacy information, including human faces, vehicles and road signs. Existing visual localization methods exploit direct visual information from these images, which may cause privacy disclosure. In this paper, a novel privacy preserving framework has been proposed to perform style-transfer on images using diffusion models and maintain semantic consistency for accurate and robust visual localization. First, it transfer the style of images using large vision models(LVMs) constrained by text prompts and semantic segmentation. Then, it estimates the camera pose by employing a feature matching algorithm for style transferred images. Experimental results on public visual localization datasets show that, in comparison with state-of-the-art methods, the proposed method can achieve comparable performance in terms of visual localization accuracy, meanwhile ensuring privacy preserving.