StanfordCars3D, CompCars3D, FGVC-Aircraft3D

本研究引入了三个大规模的3D姿态数据集：StanfordCars3D、CompCars3D和FGVC-Aircraft3D，共计包含409个细粒度类别和31,881张带有精确3D姿态标注的图像。这些数据集通过对现有细粒度识别数据集（如StanfordCars、CompCars和FGVCAircraft）进行增强，为每个子类别从ShapeNet中找到特定的3D模型，并手动调整7个连续视角参数进行标注。数据集创建过程中，利用了最先进的深度卷积神经网络（CNNs）进行姿态参数的局部搜索优化，以最大化投影掩模与分割参考之间的IoU。这些数据集主要应用于3D姿态估计研究，旨在解决细粒度对象类别从单目图像中进行3D姿态估计的问题。

This study introduces three large-scale 3D pose datasets: StanfordCars3D, CompCars3D, and FGVC-Aircraft3D, which collectively include 409 fine-grained categories and 31,881 images with accurate 3D pose annotations. These datasets are augmented from existing fine-grained recognition datasets such as StanfordCars, CompCars, and FGVCAircraft. For each subclass, specific 3D models were retrieved from ShapeNet, and annotations were conducted by manually adjusting 7 continuous viewpoint parameters. During the dataset creation process, state-of-the-art deep convolutional neural networks (CNNs) were utilized to perform local search optimization for pose parameters, so as to maximize the Intersection over Union (IoU) between the projected mask and the segmentation reference. These datasets are primarily applied to 3D pose estimation research, aiming to address the problem of 3D pose estimation of fine-grained object categories from monocular images.