Large-Displacement 3D Object Tracking with Hybrid Non-local Optimization

AbstractOptimization-based 3D object tracking is known to be precise and fast, but sensitive large inter-frame displacements. In this paper we propose a fast effective non-local method. Based on the observation that erroneous local minimum are mostly due out-of-plane rotation, hybrid approach combining optimizations for different parameters, resulting in efficient search 6D pose space. addition, precomputed robust contour-based method proposed optimization. By using long lines with multiple candidate correspondences, it can adapt frame displacements without need of coarse-to-fine search. After pre-computation, updates conducted very enabling optimization run real time. Our outperforms all previous methods both small For displacements, accuracy greatly improved (\(81.7\%\, \text {v.s.}\, 19.4\%\)). At same time, real-time speed (>50 fps) achieved only CPU. The source code available at https://github.com/cvbubbles/nonlocal-3dtracking.Keywords3D TrackingPose estimation