Synth360

我们为NeRF提出了两种非均匀射线采样方案，以适应360 ° 图像-失真感知射线采样和内容感知射线采样。我们分别使用室内和室外场景的副本和场景模型创建了一个评估数据集Synth360。在实验中，我们证明了我们的建议在准确性和效率方面成功地构建了360 ° 图像NeRF。该提案广泛适用于NeRF的高级变体。DietNeRF，AugNeRF和NeRF与所提出的技术相结合，进一步提高了性能。此外，我们证明了我们提出的方法提高了360 ° 图像中真实世界场景的质量。Synth360: https://drive.google.com/drive/folders/1suL9B7DO2no21ggiIHkH3JF3OecasQLb.We propose two non-uniform ray sampling schemes for NeRF: one tailored for 360° image distortion-aware ray sampling and the other for content-aware ray sampling. We developed the evaluation dataset Synth360 using scene replicas and their corresponding 3D models for both indoor and outdoor scenes. In our experiments, we validate that our proposed approach successfully enables the construction of high-performance 360° image NeRF models with improved accuracy and efficiency. This proposal is widely applicable to advanced NeRF variants. When integrated with DietNeRF, AugNeRF, and vanilla NeRF, the proposed technique further boosts their performance. Furthermore, we demonstrate that our proposed method improves the quality of real-world scene reconstruction from 360° images. Synth360: https://drive.google.com/drive/folders/1suL9B7DO2no21ggiIHkH3JF3OecasQLb.

搜集汇总

背景与挑战

背景概述

Synth360是一个由东京大学·国立情报学研究所于2022年发布的评估数据集，用于支持360°图像的非均匀射线采样方案研究，包含室内和室外场景的副本和模型。该数据集旨在提升NeRF在准确性和效率方面的性能，并适用于其高级变体。