DynaView: Bridging Offline Learning and Online Adaptation for UAV Viewpoint Planning

This work investigates UAV viewpoint planning for high-quality 3D reconstruction and introduces DynaView, a framework that synergizes a learned reconstructability predictor with continuous optimization. The method operates in two phases: an offline phase where a Transformer model predicts point-wise reconstructability, and an online phase where a Viewpoint Utility Learner (VUL) and Marginal Utility Maximizer (MUM) dynamically update the viewpoint set. This hybrid paradigm enables the planner to generate candidates in low-quality regions, admit or remove viewpoints based on marginal gain, and refine orientations, thereby adapting to real-time reconstruction progress while leveraging lightweight geometric descriptors and compact prediction models to keep the online optimization cost modest. Experiments on ten Urban-Scene3D buildings demonstrate that DynaView improves global reconstructability by 19.5% on average over a static baseline. These results indicate that our approach not only raises global reconstruction quality but also utilizes viewpoint resources more effectively.