OUT-OF-DISTRIBUTION RELIABILITY ESTIMATION FOR DRONE OBJECT DETECTION UNDER UNSEEN-DRONE SIMILARITY SHIFT USING SYNTHETIC DATA

Drone object detectors may perform well on known drone types while still making high-confidence errors on previously unseen drones that resemble the drones used for training. This thesis evaluates whether post-hoc out-of-distribution (OOD) scoring methods can identify potentially unreliable detectio