Shiny dataset

闪亮的文件夹包含 8 个场景，在我们的论文中使用了具有挑战性的视图相关效果。我们还在 shiny_extended 文件夹中提供了其他场景。我们论文中使用的每个场景的测试图像由按字母顺序排列的每八张图像中的一张组成。每个场景包含以下目录结构：scene/dense/cameras.bin images.bin points3D.bin project.ini images/image_name1.png image_name2.png ... image_nameN.png images_distort/ image_name1.png image_name2.png ... image_nameN .png sparse/cameras.bin images.bin points3D.bin project.ini database.db hwf_cxcy.npy planes.txtposes_bounds.npy dense/文件夹包含输入图像未失真后COLMAP的输出[1]。 images/ 文件夹包含未失真的图像。 （我们在实验中使用这些图像。） images_distort/ 文件夹包含从智能手机拍摄的原始图像。 sparse/ 文件夹包含 COLMAP 的稀疏重建输出 [1]。我们的poses_bounds.npy 类似于LLFF[2] 文件格式，稍作修改。该文件存储一个 Nx14 numpy 数组，其中 N 是摄像机的数量。该数组中的每一行都分为大小为 12 和 2 的两部分。第一部分在重新整形为 3x4 时，表示相机外部（相机到世界的转换），第二部分具有二维存储距离第一个和最后一个平面（近，远）的视角。这些距离是使用 LLFF 的代码根据场景的统计数据自动计算的。 （有关如何计算这些的详细信息，请参阅此代码）hwf_cxcy.npy 将相机内在（高度、宽度、焦距、主点 x、主点 y）存储在 1x5 numpy 数组中。 planes.txt 存储有关 MPI 平面的信息。前两个数字是从参考相机到第一个和最后一个平面（近、远）的距离。第三个数字表示平面是等距放置在深度空间 (0) 还是逆深度空间 (1) 中。最后一个数字是每个 MPI 平面的所有四个边上的填充大小（以像素为单位）。即，每个平面的总尺寸为 (H + 2 * padding, W + 2 * padding)。参考文献：[1]：运动中的 COLMAP 结构（Schönberger 和 Frahm，2016 年）。 [2]：局部光场融合：具有规定采样指南的实用视图合成（Mildenhall 等人，2019 年）。

The `shiny` folder contains 8 scenes that are utilized in our paper with challenging view-dependent effects. We also provide additional scenes in the `shiny_extended` folder. The test images for each scene used in our paper are selected as every eighth image in alphabetical order. Each scene follows the directory structure as follows:
scene/dense/cameras.bin images.bin points3D.bin project.ini images/image_name1.png image_name2.png ... image_nameN.png images_distort/ image_name1.png image_name2.png ... image_nameN.png sparse/cameras.bin images.bin points3D.bin project.ini database.db hwf_cxcy.npy planes.txt poses_bounds.npy

The `dense/` folder contains the output of COLMAP [1] after undistorting the input images. The `images/` folder contains the undistorted images, which we employ in our experiments. The `images_distort/` folder contains the raw images captured by smartphones. The `sparse/` folder contains the sparse reconstruction output of COLMAP [1].

Our `poses_bounds.npy` follows a slightly modified version of the LLFF [2] file format. This file stores an N×14 numpy array, where N denotes the total number of cameras. Each row in this array is split into two segments with sizes 12 and 2, respectively. When reshaped into a 3×4 matrix, the first segment represents the camera extrinsic parameters (camera-to-world transformation), while the second segment stores the distances to the near and far planes in 2D format. These distances are automatically calculated based on the scene statistics using the official LLFF code; please refer to this code for detailed instructions on the distance calculation process.

The `hwf_cxcy.npy` file stores the camera intrinsic parameters (height, width, focal length, principal point x-coordinate, principal point y-coordinate) in a 1×5 numpy array.

The `planes.txt` file stores information related to MPI planes. The first two numerical values are the distances from the reference camera to the near and far planes. The third value indicates whether the planes are arranged equidistantly in the linear depth space (0) or inverse depth space (1). The last value is the padding size (in pixels) applied to all four edges of each MPI plane. Namely, the total dimensions of each plane are (H + 2 * padding, W + 2 * padding).

References:
[1] COLMAP: Structure from Motion (Schönberger and Frahm, 2016)
[2] Local Light Field Fusion: Practical View Synthesis with Prescriptive Sampling Guidelines (Mildenhall et al., 2019)