Data-driven methods for non-visual indoor localization and extreme pose estimation

This thesis addresses the problems of non-visual localization and extreme pose estimation in indoor environments by adopting novel data-driven methods. Robust indoor localization allows building mobile location-aware services; while pose estimation of non-overlapping indoor panoramas brings about extensive applications in real-estate by granting easy data acquisition. We infer an indoor location of history by fusing IMU, WiFi sparse positions, and a floorplan image. This multi-modal fusion employs inertial navigation, non-linear optimization, and CNN. Our results are twice as accurate and a few orders of magnitude denser than the current standard, while maintaining energy efficient. We solve pose estimation of house panoramas with little to no overlaps by inferring room layouts and annotations, and utilizing CNN and Transformer modules. Our system's accuracy is competitive with SOTA, yet much faster and capable of handling more complex house structures. We will publicly share codes and models.