Motion-Regime-Aware Feature Decoupling for Transformer-Based Monocular Camera Relocalization

Monocular camera relocalization recovers a six-degree-of-freedom pose from one RGB image, but direct absolute pose regression typically predicts translation and rotation from one terminal representation. We propose Decoupled SwinPose, a hierarchical Swin-Tiny regressor that instead reads translation from a shallow, higher-resolution Stage 1 map and rotation from the deep, contextual Stage 3 representation, testing this asymmetric-readout hypothesis through three falsifiable predictions. Across three TUM RGB-D motion regimes and all seven Microsoft 7-Scenes environments, against constant-pose, retrieval, and matched shared-terminal controls, it achieves the lowest three-seed mean translation and rotation error on all ten evaluated sequences within the matched reimplemented cohort. Relative to the strongest alternative learned model within this cohort, translation error falls by up to 29.6% on TUM RGB-D and 52.6% on 7-Scenes, and rotation error by up to 27.2% and 54.4%, respectively. Routing ablations, including a capacity-matched control, support T1-R3 as the strongest overall trade-off among six tested configurations, and matched profiling shows the readout adds only 0.29% parameters with essentially identical FLOPs and FP32 cost relative to a shared-terminal baseline on an NVIDIA L4. These results support asymmetric multilevel readouts as an effective, low-cost architectural prior for transformer-based monocular pose regression in the evaluated indoor setting.