本文提出了一个可以表示大规模场景的 NeRF ,它将大场景分解为单独训练的 NeRF ,这使得渲染时间与场景大小无关。给每个NeRF加入了appearance embeddings, learned pose refinement, and controllable exposure,并且引入了NeRF之间的对齐模块。作者用280万张图像重建了旧金山的一个街区,并且图片有准确的相机姿态和位置数据。
【参考】 Tancik M, Casser V, Yan X, et al. Block-nerf: Scalable large scene neural view synthesis[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022: 8248-8258.
渲染城市级别的大场景时需要把场景解耦多个独立的NeRF训练,这种解耦使得场景的尺寸和渲染的时间解耦,能够渲染任意大的场景,并且允许每个block更新环境。对每个nerf加了appearance embeddings,learned pose refinement,controllable exposure,对相邻nerf之间加了对齐的步骤使其可以无缝衔接,实现了280万张图片训练出旧金山的一整个...
主页:http://waymo.com/research/block-nerf Block-NeRF是一种通过使用多个紧凑的nerf(每个nerf都适合内存)来表示环境,从而实现大规模场景重建的方法。在推断时,Block-NeRF无缝结合给定区域的相关nerf的渲染。论文中,使用3个多月收集的数据(2.8w)重建了旧金山的Alamo广场社区。Block-NeRF可以更新环境的单个块,而无需...
Block-NeRF: Scalable Large Scene Neural View Synthesis Matthew Tancik1⇤ Vincent Casser2 Xinchen Yan2 Sabeek Pradhan2 Ben P. Mildenhall3 Pratul Srinivasan3 Jonathan T. Barron3 Henrik Kretzschmar2 1UC Berkeley 2Waymo 3Google Research Alamo Square, SF June Sept. Block-NeRF 1 ...
We present Block-NeRF, a variant of Neural Radiance Fields that can represent large-scale environments. Specifically, we demonstrate that when scaling NeRF to render city-scale scenes spanning multiple blocks, it is vital to decompose the scene into individually trained NeRFs. This decomposition decou...
Block-NeRF: Scalable large scene neural view synthesis. arXiv, 2022. 3 [35] Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P. Srinivasan, Jonathan T. Barron, and Ren Ng. Learned initializations for optimizing coordinate-based neural...
Large, heavy objects with distinct signatures such as the tea box have high detection accuracy across different numbers of inputs (N). i, Original first-layer convolution filters (3 × 3) learned by the network shown in Fig. 2a for N = 1 inputs. j, Visualization of the first-layer ...
• The core of NICE-SLAM is a hierarchical, grid-based neural implicit encoding. In contrast to global neural scene encodings, this representation allows for local up- dates, which is a prerequisite for large-scale approaches. • We conduct extensive evaluations on various datasets...
By contrast with recent patch-based methods, we rely on a "holistic" approach: We apply to the detected objects a Convolutional Neural Network (CNN) trained to predict their 3D poses in the form of 2D projections of the cor- ners of their 3D bounding boxes. This, however, is not suf-...