Middlebury 2003

这些数据集由 Daniel Scharstein、Alexander Vandenberg-Rodes 和 Rick Szeliski 创建。它们由具有复杂几何形状的高分辨率立体序列和像素精确的地面实况视差数据组成。地面实况差异是使用一种新技术获得的，该技术采用结构化照明并且不需要校准投光器。有关详细信息，请参阅我们的 CVPR 2003 论文。_x000D_
我们的新数据集“Cones”和“Teddy”的四分之一大小 (450 x 375) 版本可在下方下载。每个数据集包含 9 个彩色图像（im0..im8）和 2 个视差图（disp2 和 disp6）。这 9 幅彩色图像形成了一个多基线立体序列，即它们是从沿 x 轴从左到右的等距视点拍摄的。图像被校正，因此所有图像运动都是纯水平的。要测试双视图立体算法，应使用两个参考视图 im2（左）和 im6（右）。为这两个视图提供了具有四分之一像素精度的地面实况差异。对于灰度级 1 .. 255，视差使用比例因子 4 进行编码，而灰度级 0 表示“未知视差”。因此，编码的视差范围为 0.25 .. 63.75 像素。

These datasets were created by Daniel Scharstein, Alexander Vandenberg-Rodes, and Rick Szeliski. They consist of high-resolution stereo sequences with complex geometric shapes and pixel-accurate ground-truth disparity data. The ground-truth disparities were obtained using a novel technique that employs structured lighting and does not require projector calibration. For more details, please refer to our CVPR 2003 paper.
Quarter-size (450 × 375) versions of our new datasets "Cones" and "Teddy" are available for download below. Each dataset contains 9 color images (im0..im8) and two disparity maps (disp2 and disp6). The 9 color images form a multi-baseline stereo sequence, captured from equally spaced viewpoints along the x-axis from left to right. The images are rectified such that all image motion is purely horizontal. To test two-view stereo algorithms, the two reference views im2 (left) and im6 (right) should be used. Ground-truth disparities with quarter-pixel accuracy are provided for these two views. Disparities are encoded with a scale factor of 4 for gray levels 1 through 255, while gray level 0 indicates "unknown disparity". As a result, the encoded disparities range from 0.25 to 63.75 pixels.

搜集汇总

背景与挑战

背景概述

Middlebury 2003数据集由Daniel Scharstein等人创建，提供高分辨率立体序列和像素级真实视差数据，适用于立体视觉算法评估。该数据集包含'Cones'和'Teddy'的450x375版本，具有9个彩色图像和2个视差图，视差编码精度为四分之一像素，由微软·明德大学于2003年发布。