Research on CAD Model Construction for Indoor Scenes Based on Dynamic View Planning

We propose an active RGB-D framework that bridges active perception with task-oriented, manipulation-ready scene modeling in a single pass. At its core is a closed-loop pipeline integrating three key components: (1) a robust fusion of geometric and semantic information at the instance level; (2) a novel planning strategy, featuring a Next-Best-Object (NBV) module that guides an adaptive NBV planner by prioritizing objects based on visual saliency, information gain, and reachability; and (3) functionally equivalent CAD model replacement. By focusing sensing efforts on the most informative targets, our system efficiently generates semantically grounded, interaction-ready CAD scenes. Evaluated on SUNCG and SceneNN, it achieves 77% and 61% average precision (AP), respectively, demonstrating robust generalization across synthetic and real-world indoor environments. The key novelty is the system-level, single-pass design that closes the perception-to-interaction loop, enabling autonomous scene understanding directly applicable to robotic manipulation.