PS-Wild

我们还为评估目的创建了一个测试数据集。生成图像的计算管道与训练数据集的计算管道相同，但是使用不同的3-D资产进行公平评估; 来自CGTrader的25个对象，来自shareextures的50个材料和来自sIBL存档的50个环境地图。对于每个3-D模型，我们分配两组材质和环境图，从而产生50组不同的对象，材质和环境图。与训练数据集不同，我们会根据每个纹理类别仔细选择六个纹理; 在sharretextures中分类的混凝土，织物，地板，地面，木材和金属。为了正确评估一种方法在各种照明条件下的性能，我们使用三种不同的照明方法为同一组对象和材料渲染图像; (a) 单方向照明 (均匀采样)，(b) HDRI照明 (与培训相同) 和 (c) (a) 和 (b) 的混合物。图像分辨率也是512 × 512，但是图像数量是32，以评估输入图像数量不同的性能。

We also developed a test dataset for evaluation purposes. The computational pipeline for image generation is identical to that used for the training dataset, yet we employ distinct 3D assets to guarantee a fair evaluation: 25 objects obtained from CGTrader, 50 materials sourced from shareextures, and 50 environment maps from the sIBL Archive. For each 3D model, we assign two sets of materials and environment maps, yielding 50 unique combinations of objects, materials, and environment maps. Unlike the training dataset, we carefully select six textures that fall into six predefined categories in shareextures: concrete, fabric, flooring, ground, wood, and metal. To accurately evaluate a method's performance across diverse lighting conditions, we render images for the same set of objects and materials using three distinct lighting strategies: (a) unidirectional lighting (uniform sampling), (b) HDRI lighting (consistent with that used in the training dataset), and (c) a hybrid combination of (a) and (b). The image resolution is fixed at 512 × 512, and we generate a total of 32 images to evaluate performance under varying numbers of input images.

搜集汇总

背景与挑战

背景概述

PS-Wild是一个用于光度立体网络评估的测试数据集，包含多种3-D资产生成的图像，提供三种照明方法渲染的图像，用于评估不同照明条件下的性能。