One Image, One Model: Semantic-Driven Single- Image 3D Reconstruction

ackground: Generating high-fidelity 3D models from a single 2D image remains a persistent challenge in computer vision. Traditional approaches rely heavily on photogrammetry or multi-view reconstruction, making single-image reconstruction difficult in scenarios where such data is unavailable. To address this limitation, we propose a framework that enables single-image 3D reconstruction while complementing rather than replacing photogrammetric techniques. Methods: Our approach integrates multiple AI-driven techniques for structured 3D reconstruction. The pipeline begins with segmentation using the Segment Anything Model (SAM), ensuring precise object isolation. Object classification follows, employing the KnowDIP knowledge-driven pipeline, which combines ontology-based reasoning, rule-based algorithms, and feature extraction to enhance classification accuracy. Linked Open Data (LOD) sources, such as DBpedia and Wikidata, provide geometric constraints and contextual knowledge, guiding the reconstruction process. These constraints refine the output of the Large Reconstruction Model (LRM), which generates an initial single-view 3D mesh. The final stage applies mesh repair and semantic refinement, ensuring the reconstructed model is manifold and suitable for visualization, holography, and 3D printing. Results: We evaluated the framework on 100 diverse objects, achieving centimeter-level accuracy in reconstruction. Our method successfully generates high-fidelity 3D models, particularly in cases where multi-view photogrammetric approaches are impractical. However, results indicate a slight bias toward orthogonal perspectives, as the reconstruction process tends to favor the dominant input orientation, occasionally leading to minor geometric inaccuracies in cases of extreme perspective distortions. Conclusion: Our framework provides an effective alternative for single-image 3D reconstruction in cases where traditional multi-view approaches are infeasible. By combining segmentation, classification, ontology-based reasoning, and geometry-aware reconstruction, our method advances the field of AI-driven 3D modeling. This approach is deployed at shinai3d.com, allowing free testing of the method. Future work will focus on refining perspective correction techniques to further improve geometric accuracy in challenging cases.