Sparse-View 3D Reconstruction of Unstructured Objects in Confined Environments

Reconstructing unstructured objects, such as flexible cables in confined industrial environments, faces challenges of geometric breakage and floating artifacts. These issues, caused by weak textures, strong reflections, and sparse views, hinder collision avoidance in digital twin applications. To address them, an improved 3D Gaussian Splatting (3DGS) framework integrating Multi-Resolution Merging Network (MRMN) and Floating Pruning (FP) is proposed. First, MRMN extracts depth priors with topological continuity by cascading low-resolution global skeletons and high-resolution edge features. Second, a depth consistency loss function based on Pearson Correlation Coefficient (PCC) is designed to constrain geometric convergence. Finally, an FP technique based on Dip Test modality analysis is introduced to adaptively remove false obstacles. Experimental results indicate that the proposed method effectively restores the continuous topology of black cables with diameters of 2 ~ 5 mm. Compared with traditional methods focusing on visual metrics, this study adopts a geometry-first strategy. It significantly improves the physical purity and connectivity of the model, thereby eliminating potential collision risks in industrial robot path planning.