EIF-6-D: Enhancing Category-Level 6-D Pose Estimation Through Dual-Stream Implicit-Explicit Transformation

Category-level 6D pose estimation, predicting the 3D position, orientation, and size of unseen objects, is pivotal for applications in robotics and augmented reality. Existing methodologies typically fall into two categories: shape prior-based approaches, which establish correspondences through explicit deformation fields but are hindered by intra-class shape variability, and shape prior-free frameworks, which directly regress observation features but suffer from robustness issues under occlusion. To address these challenges, we introduce EIF-6D, a dual-stream transformation network that integrates explicit and implicit deformation mechanisms for precise alignment between shape priors and observed objects. Through a Prior Transformation Module (PTM) and a Prior Feature Transformation Module (PFTM), EIF-6D achieves coarse-to-fine optimization, enabling high-precision pose estimation. Experiments on the NOCS-REAL275 dataset demonstrate that EIF-6D significantly outperforms existing methods, achieving 50.6% and 57.4% average precision under the stringent 5°2cm and 5°5cm metrics, respectively. Our approach sets a new state-of-the-art performance among shape prior-based methods. Codes are released at https://github.com/YuanChenWe/EIF6D.