ThermalManhattan: Robust Thermal Compass under Radiometric Flatness in Manhattan World

We present a drift-free, matching-free 3-DoF visual compass for monocular thermal imagery. Thermal cameras sense emitted radiation and operate regardless of illumination, but when surfaces reach thermal equilibrium, their texture vanishes, so point features collapse and the genuine, shape-preserving cues survive only as structural lines at object boundaries. Existing line-based thermal approaches still match lines across frames, where lines are not extracted consistently from frame to frame and tracking further degrades as the lines become too sparse. We instead avoid inter-frame line matching and, to the best of our knowledge, are the first to bring the Manhattan world (MW) assumption to thermal imagery, estimating the absolute camera orientation independently for every frame. Given a gravity vector from the drone, we recover the remaining heading angle (yaw) from structural lines, so each frame is fixed on its own and no drift accumulates across NUC drops or aggressive motion. We collect six outdoor man-made UAV sequences and, since structure-from-motion is unreliable on monocular thermal, build pseudo ground-truth (GT) trajectories with RTK-GPS priors. We publicly release this benchmark, the first paired RGB–thermal night-time UAV dataset under the MW assumption. On this benchmark, the proposed method outperforms existing rotation estimators and stays stable over 20+ rotational loops on which the onboard gyro integrates 15° of yaw drift at the end.