An Improved Algorithm for Object Pose Estimation Via 3D Gaussian Splatting

Object pose estimation, a pivotal task in computer vision, has significant applications in domains such as industrial automation and augmented reality. To address the limitations of traditional 3D reconstruction algorithms, including their imbalance between computational efficiency and accuracy, as well as sensitivity to initial poses, this study proposes a pose estimation framework based on enhanced 3D Gaussian splatting. The framework introduces three key components: a selector module that optimizes initial pose generation by evaluating image similarity, a dynamic loss combination strategy integrating matching loss, perceptual loss, and pixel comparison loss to improve convergence, and an improved 3D Gaussian model for enhanced reconstruction accuracy. Experimental results on the Realistic Synthetic 360° and Mip-NeRF 360 datasets demonstrate the method's robustness, with improvements in both success rate and accuracy compared to the iComMa baseline.