DSPose: Decoupled Real-Time Pose Estimation From Semantic-RGB Radiance Rays

Six-degree-of-freedom (6-DoF) pose estimation receives substantial attention, particularly in embodied agents and augmented reality. Recent methods explore a ray-selection paradigm, which reframes pose estimation from slow "analysis-by-synthesis" to direct ray selection. However, this paradigm still struggles with the selection ambiguity stemming from the strong dependence on RGB feature, and also with the unavoidable coupling problem of translation and rotation during optimization, which inherently leads to translation-rotation degeneracy. Hence, we introduce a robust selection approach DSPose that first generates Semantic-RGB radiance rays to ensure their quality and then stabilizes the selection through a decoupled optimization of translation and rotation. To achieve this, we first propose the Geometric-Resampling Semantic Representation (GRSR), which forges a geometrically accurate scene structure. This representation enhances the discriminability hugely to resolve selection ambiguities by embedding each ray with rich semantic information. Given the inherent entanglement of translation and rotation in a coupled selection process, we propose the Euclidean-Lie Decoupled (ELD) mechanism to ensure stability. Thereby, DSPose enables robust and efficient selection by ensuring both candidate-ray discriminability and selection stability. Extensive experiments show DSPose establishes a new state-of-the-art (SOTA), achieving unprecedented accuracy and real-time performance to make robust, prior-free pose estimation practical. To the best of our knowledge, this is the first work to drastically reduce both angular error and translation error in real-time, breaking the long-standing performance barrier for prior-free methods.