A comprehensive Gaussian splatting evaluation system: from geometric consistency to novel view synthesis

This paper proposes a comprehensive evaluation system for Gaussian Splatting (GS) that addresses both the Structure from Motion (SfM) and GS model processes, covering a full range of assessment methods from geometric consistency to Novel View Synthesis (NVS). Firstly, to address the prior neglect of SFM evaluation in 3D reconstruction works, the system integrates evaluation metrics applicable to both SFM and GS models. Secondly, considering the scale discrepancies between reconstructed and real-world targets, it implements automated scale alignment and point cloud registration to avoid the limitations of manual scale alignment. Subsequently, given that current evaluations of novel view synthesis emphasize information restoration and texture reconstruction, this paper constructs a geometric evaluation metric based on normal vector information to facilitate quantitative analysis of reconstruction structures. Subsequent experiments demonstrate that this new metric exhibits higher sensitivity and broader applicability than existing evaluation measures. Finally, the paper introduces a new evaluation dataset, supporting algorithm degradation research for 3D GS in edge regions and close-range observations. Experiments confirm that the proposed system serves as an effective assessment tool for GS techniques, validating that the new dataset contributes significantly to performance exploration in current 3D GS applications.