Monocular 3D Point Cloud Reconstruction Dataset (Mono3DPCL)

As with most AI methods, a 3D deep neural network needs to be trained to properly interpret its input data. More specifically, training a network for monocular 3D point cloud reconstruction requires a large set of recognized high-quality data which can be challenging to obtain. Hence, this dataset contains the image of a known object alongside its corresponding 3D point cloud representation. To collect a large number of categorized 3D objects, we use the ShapeNetCore (https://shapenet.org) dataset. It is a densely annotated subset of the ShapeNet dataset comprising 55 common object categories with 51,300 unique synthetic 3D models. These 3D models, however, were in mesh format and needed to be converted to the ground truth 3D point cloud. To convert these 3D mesh models to point clouds and to capture a single image from them, we used the Open3D library (https://www.open3d.org). The objects were positioned at the center of the scene and an image of each object was exported using a fixed-view camera for all objects. This way, we collectively prepared ~16000 data pairs in 8 groups of datasets in categories of the airplane, bottle, car, bench, sofa, cellphone, bike, and speaker. We dedicated a subdivision of the data in each category to the training (85%), validation (5%), and test (10%) phases.

与大多数人工智能方法类似，3D深度神经网络需经过训练以准确解读其输入数据。具体而言，训练用于单目3D点云重建的网络，需要大量被认可的高质量数据集，而此类数据集的获取颇具挑战性。因此，本数据集包含已知物体的图像及其相应的3D点云表示。为收集大量分类化的3D物体，我们采用了ShapeNetCore（https://shapenet.org）数据集。它是ShapeNet数据集的一个密集标注子集，包含55种常见物体类别和51,300个独特的合成3D模型。然而，这些3D模型以网格格式存在，需要转换为真实3D点云。为了将这些3D网格模型转换为点云，并从中捕获单幅图像，我们使用了Open3D库（https://www.open3d.org）。物体被放置在场景的中心，并使用固定视角的相机为所有物体导出图像。通过此方法，我们共准备了约16000个数据对，分为8个数据集组，涵盖了飞机、瓶子、汽车、长椅、沙发、手机、自行车和扬声器等类别。在每个类别中，我们将数据集分为训练（85%）、验证（5%）和测试（10%）三个阶段。