Megascans Plants

现代3d感知图像合成基准有两个问题: 1) 它们包含非常
类似的全局几何形状 (例如，人或猫的脸，汽车和椅子)，以及2) 它们的相机较差
覆盖范围。此外，当对象特征 (例如，FFHQ) 时，其中一些 (例如，FFHQ) 包含3d偏差 (例如，
微笑概率、凝视方向、姿势或发型) 与相机位置相关 [6]。作为
结果，这不允许评估模型表示基础几何的能力，并且
使人们更难理解性能是来自方法上的变化还是更好
数据预处理。
为了缓解这些问题，我们引入了两个新的数据集: Megascans植物 (M-Plants) 和Megascans
食物 (M-食物)。为了构建它们，我们从Quixel Megascans6获得 ≈ 1，500模型
来自植物，
蘑菇和食品类别。Megascans是对真实物体的高质量扫描
几乎与真实没有区别。对于蘑菇和植物，我们将它们合并到相同的食物中
类别，因为他们自己的模型太少了。
我们用相机在搅拌机中渲染所有模型，随机均匀分布在
半径3.5的球体和 π/4的视场。在渲染时，我们将每个模型缩放为 [− 1,1] 3
立方体并丢弃那些尺寸小于2的模型。我们呈现128视图
从固定距离到对象中心的每个对象，从整个对象上的均匀采样点
球体 (甚至从下面)。对于M-Plants，我们另外删除了那些少于
平均0.03像素强度 (计算为像素和视图上的平均 α 值)。这是
需要去除小草或树叶，这将占用太少的像素。作为一个
结果，该程序产生了植物类别的1,108模型和食品的199模型
类。

Modern 3D-aware image synthesis benchmarks suffer from two critical limitations: 1) They contain highly similar global geometries (e.g., human or cat faces, cars, and chairs), and 2) Their camera coverage is insufficient. Furthermore, some benchmarks (e.g., FFHQ) exhibit 3D biases, where object attributes such as smile probability, gaze direction, pose, or hairstyle are correlated with camera pose [6]. As a result, these benchmarks fail to enable evaluation of a model's ability to represent underlying geometry, and make it more difficult to discern whether performance stems from methodological improvements or better data preprocessing.

To address these issues, we introduce two novel datasets: Megascans Plants (M-Plants) and Megascans Food (M-Food). To construct these datasets, we obtained approximately 1,500 3D models from the plant, mushroom, and food categories via Quixel Megascans [6]. Megascans comprises high-quality scans of real-world objects that are nearly indistinguishable from their physical counterparts. For mushrooms and plants, we grouped them into the same food category due to the insufficient number of individual models in their respective categories.

We rendered all models using cameras in Blender, with camera positions uniformly and randomly sampled on a sphere of radius 3.5, and a field of view set to π/4. During rendering, we scaled each model to fit within a [-1, 1]^3 bounding cube, and discarded models with a dimension smaller than 2. We rendered 128 views for each object, with the camera placed at a fixed distance from the object's center, sampling positions uniformly across the entire sphere (including below the object). For the M-Plants dataset, we additionally removed models with an average pixel intensity lower than 0.03, calculated as the average alpha value across all pixels and views. This step is necessary to eliminate tiny grass or leaves that occupy an insufficient number of pixels in the rendered images. As a result, this pipeline yields 1,108 models for the plant category and 199 models for the food category.

搜集汇总

背景与挑战

背景概述

该数据集旨在解决现有3D感知图像合成基准中几何形状相似和相机覆盖不足的问题，通过从Quixel Megascans获取高质量真实物体扫描构建而成。它包含植物、蘑菇和食品类别，其中植物和蘑菇被合并，并经过Blender渲染、相机均匀采样及模型筛选处理，最终生成了1,108个植物类别模型。