UNITE-NBV: Uncertainty-Driven and Information-Enhanced Gain Estimation for Next Best View

Abstract

Next Best View (NBV) algorithms are a critical area of research in 3D reconstruction. They aim to efficiently reconstruct 3D scenes by maximizing information gain from the next optimal viewpoint. However, current NBV methods often neglect the importance of high-quality candidate view sampling, leading to inconsistent quality of the candidate viewpoint set. Moreover, these methods frequently encounter difficulties in extracting effective information for accurate information gain estimation, especially when reconstructing complex objects or large-scale scenes. To address these challenges, we propose UNITE-NBV. Our method achieves accurate next best view selection across various scenes by effectively uniting the proposed high-quality candidate view sampling strategy and information gain estimation network. Specifically, we introduce Spherical Uncertainty Sampling (SUS), a novel candidate viewpoint sampling method. This method calculates the scene's uncertainty field and maps it into a spherical sampling space. Within this space, the uncertainty distribution guides the sampling of high-quality candidate viewpoints. Additionally, we propose a Multi-Expert Information Gain Network (MEIGN) that performs disentangled encoding and expert encoder sequence processing on features extracted from the reconstructed scene. These processed features are then dynamically fused using sparse gating and Multi-Head Self-Attention, enabling accurate information gain estimation for candidate views. Extensive experimental results on both the small-object dataset ShapeNet and various large-scale 3D scene datasets demonstrate the effectiveness and superior performance of our proposed method. The code will be released.