Accurate one-shot object pose estimation with Gaussian Splatting reconstruction

We propose a novel and flexible framework for instance-based 6D pose estimation from a single target image. Unlike existing methods, our approach reconstructs high-fidelity and dense 3D object models from reference images using Gaussian Splatting, providing richer geometric and appearance cues. To explicitly bridge the 2D target image with the 3D canonical model, we render a set of virtual viewpoints around the reconstructed 3DGS model, generating multiple RGB–depth pairs to capture diverse texture and geometric information. Robust feature matching establishes 2D–2D correspondences between the virtual views and the target image, which are back-projected to produce accurate 2D–3D correspondences in the canonical coordinate system. An initial 6-DoF pose is computed via the PnP algorithm and subsequently refined using a 3DGS-based pose optimizer. Experimental results demonstrate that our method achieves superior accuracy compared to state-of-the-art baselines, highlighting the effectiveness of integrating dense 3DGS reconstruction with virtual-view correspondence learning for precise 6D object pose estimation.