Advanced Monocular 6D Pose Estimation with Enhanced 3D Coordinate Maps via Feature Fusion and Deformable Decoupled Regression

Abstract

Monocular 6D pose estimation remains challenging due to low-quality 3D coordinate maps caused by occlusion, texture-less surfaces, and spatial detail loss in encoder-decoder networks. This paper presents an efficient monocular framework that improves pose accuracy by enhancing the quality of dense 3D coordinate maps via feature fusion and deformable decoupled regression. We propose a lightweight feature fusion module composed of long-range-aware feature fusion and multiscale feature fusion, which integrates spatial details and longrange dependencies using deformable convolutions. Meanwhile, a disentangled pose regression network is designed to separately estimate rotation and translation for better optimization. Extensive experiments on Linemod and Linemod-Occluded datasets show that our method achieves competitive performance with state-of-the-art dense correspondence methods while maintaining real-time speed. Ablation studies verify the effectiveness of each proposed component, and the feature fusion module generalizes favorably to other representative pose solvers. Finally, our method provides an efficient and robust solution for monocular 6D pose estimation, with broad application prospects in real-world robotic scenarios.