Current view planning (VP) systems usually adopt an iterative pipeline with next-best-view (NBV) methods that can autonomously perform 3D reconstruction of unknown objects. However, they are slowed down by local path planning, which is improved by our previously proposed set-covering-based network SCVP using one-shot view planning and global path planning. In this work, we propose a combined pipeline that selects a few NBVs before activating the network to improve model completeness. However, this pipeline will result in more views than expected because the SCVP has not been trained from multiview scenarios. To reduce the overall number of views and paths required, we propose a multiview-activated architecture MA-SCVP and an efficient dataset sampling method for view planning based on a long-tail distribution. Ablation studies confirm the optimal network architecture, the sampling method and the number of samples, the NBV method and the number of NBVs in our combined pipeline. Comparative experiments support the claim that our system achieves faster and more complete reconstruction than state-of-the-art systems. For the reference of the community, we make the source codes public.
翻译：当前的视角规划系统通常采用迭代流程和自主进行 3D 重建的下一步最佳视角 （NBV）方法。然而，它们因需要进行局部路径规划而变慢，而我们之前提出的基于集合覆盖网络的视角规划系统 SCVP 利用一次性视角规划和全局路径规划进行了改进。在这项工作中，我们提出了一种组合流程，先选择几个 NBV，然后再激活网络来提高模型的完整性。然而，该流程会比预期需要更多的视角，因为 SCVP 没有从多视角情况下进行训练。为了减少所需的视角和路径数量，我们提出了一种多视角激活体系结构 MA-SCVP 和一种基于长尾分布的高效数据集采样方法。削减实验证实了最优的网络体系结构、采样方法和采样数量、NBV 方法和 NBV 数量在我们的组合流程中的表现。比较实验支持我们的系统比现有系统实现更快、更完整的重建。为了让学术界可以参考，我们公开了源代码。
相关内容
微信扫码咨询专知VIP会员