We develop LoTIS, a transformer that localizes a query image relative to an image trajectory (direction, distance and uncertainty to every trajectory frame), and use it to build navigable topological place graphs of buildings from ordinary RGB video, without metric priors or per-scene tuning. The project trains and evaluates model variants (from-scratch ablations, fine-tuning of existing checkpoints, new geometric training objectives) on precomputed DINOv3 features from Habitat, TartanAir, TartanGround and real-world datasets, and evaluates graph quality (loop-closure recall and precision, layout error) on held-out buildings. The PI is a doctoral student at the Division of Robotics, Perception and Learning (RPL), KTH Royal Institute of Technology, funded by the Wallenberg AI, Autonomous Systems and Software Program (WASP), Knut and Alice Wallenberg Foundation. Main supervisor: Olov Andersson, RPL, KTH Royal Institute of Technology.