TL;DR通过分类的方法,提出了一种下一个最佳视角选择的方案,着重于通过一对被动获取的 2D 视图获得最佳可能的 3D 重建质量,并通过地面真实体素的间接 3D 监督来联合训练分类器和重构网络。
Abstract
We present a classification based approach for the next best view selection
and show how we can plausibly obtain a supervisory signal for this task. The
proposed approach is end-to-end trainable and aims to get t