Sparse3DPR: Training-Free 3D Hierarchical Scene Parsing and Task-Adaptive Subgraph Reasoning from Sparse RGB Views

The introduction of Sparse3DPR marks a significant advancement in the field of 3D scene understanding, leveraging the strengths of large language models without the need for extensive training. By utilizing sparse RGB inputs and a hierarchical plane-enhanced scene graph, Sparse3DPR improves reasoning clarity and accuracy. Experimental results indicate a remarkable 28.7% improvement in EM@1 and a 78.2% speedup over ConceptGraphs on the Space3D-Bench, showcasing its efficiency. Additionally, its performance on ScanQA demonstrates comparable results to traditional training-based methods, further validating its robustness and generalization capabilities. This framework not only enhances the efficiency of 3D scene reasoning but also opens up new possibilities for applications in various domains, making it a noteworthy development in artificial intelligence.