cudar/ImageNet-PD

ImageNet-PD是一个用于评估模型对透视畸变鲁棒性的新基准数据集。透视畸变会导致图像中视觉概念的形状、大小、方向、角度和其他空间关系发生前所未有的变化。该数据集基于ImageNet验证集，通过在不同方向上合成畸变而创建，包含八个子集：四个对应不同方向（左、右、上、下）并具有黑色背景（PD-L、PD-R、PD-T、PD-B），另外四个具有相同方向但使用边界像素进行集成填充背景（PD-LI、PD-RI、PD-TI、PD-BI）。透视畸变由相机位置、深度、内在参数（如焦距和镜头畸变）以及外在参数（如旋转和平移）等因素引起，这些因素影响3D场景到2D平面的投影，对语义解释和局部几何造成挑战。该数据集旨在帮助开发更鲁棒的计算机视觉方法。ImageNet-PD is a new benchmark dataset for robustness evaluation against unprecedented changes in shape, size, orientation, angles, and other spatial relationships of visual concepts in images due to perspective distortion. It is derived from the ImageNet validation set by synthesizing distortions in different orientations and consists of eight subsets: four corresponding to four orientations (left, right, top, bottom) with black background (PD-L, PD-R, PD-T, PD-B), and the other four with the same orientations but with integrated padding background using boundary pixels (PD-LI, PD-RI, PD-TI, PD-BI). Perspective distortion results from factors such as camera positioning, depth, intrinsic parameters like focal length and lens distortion, and extrinsic parameters like rotation and translation, which impact the projection of 3D scenes onto 2D planes and pose challenges for semantic interpretation and local geometry. The dataset is designed to facilitate the development of robust computer vision methods.