(BMVC 2025 Oral) MonoTracker: Monocular RGB-Only 6D Tracking of Unknown Object

Estimating the six degrees of freedom (6D) pose of unknown objects using only monocular RGB images is a challenging task, especially when dealing with textureless and small objects. In this paper, we propose a novel pipeline, MonoTracker, for 6D object pose estimation and tracking that operates without any prior depth information. MonoTracker is a model-free, RGB-only, 6D detector that works on unseen objects. It leverages state-of-the-art pre-trained deep learning models, enabling zero-shot 6D pose estimation by jointly optimizing object poses and correcting scale inconsistencies in monocular depth predictions. We validate our method on three public datasets – YCBInEOAT, HO3D, and BEHAVE – demonstrating significant improvements over the state of the art. As a downstream application, we also show that the estimated camera poses can be used as input in NeRF pipelines, facilitating novel-view synthesis. Our results highlight the potential of monocular RGB inputs for accurate 6D object tracking and reconstruction in real-world scenarios. The code will be made public.

In The Thirty Sixth British Machine Vision Conference