DensePoint: Densely Annotated 3D Shape Database
Abstract
A point cloud is an important representation of 3D objects or real world scenes and plays an important role in various applications such as autonomous driving and virtual reality. However, applying a data-driven approach to a point cloud has still not been studied well compared to the boom of deep learning in image-related tasks. We think that one of the reasons is the lack of a well-annotated point cloud database. In this paper, we propose DensePoint, which is a densely sampled and annotated point cloud database containing over 10,000 instances. Each point cloud contains 40,000 points, and each point is associated with two sorts of information: RGB value and part annotation. This dataset is a promising candidate for various tasks. We demonstrate one example in which a point cloud is colorized in a data-driven way by utilizing Generative Adversarial Networks (GANs). The results show that our generated coloured point clouds have a reasonably good colour pattern based on different parts.
Fig 1.One example of each category from our DensePoint dataset. Leftmost image of each pair is represented by RGB value, and rightmost image is same point cloud represented by part label.
Data
Fig 2. Statistics of DensePoint.
DensePoint a point cloud dataset extended from ShapeNet and ShapeNetPart, with each point in a point cloud associated with 4 types of features:
1. XYZ coordinates
2. RGB color
3. normal vector
4. part label
Following ShapeNetPart, there are 50 parts among 16 categories.
We save point clouds in ply format with binary_little_endian 1.0 encoding.
For each category, we prepare two folders, ply_files and img.
The former contains all ply format files of point clouds, and the latter contains all visualization image pairs of point clouds, with the left one visualizing a point cloud based on its color and the right one based on different parts.
We also provide train/val/test split following ShapeNet official settings in json format files.
You could download DensePoint by clicking following button.(~6.5 Gb)
Point Cloud Colorization
Fig 3.Our generative adversarial network’s architecture. Generator, modified from PointNet segmentation network, tries to predict point-wise colour for N x 3 input point cloud without colour. Discriminator, modified from PointNet classification network, takes two batches of point clouds as input and tries to distinguish them. One batch is generated colour concatenated with point cloud along feature axis, and the other batch is real coloured point cloud.
Fig 4.Colorization results on test dataset. Leftmost image in pair is ground-truth coloured point cloud while rightmost image in pair is colorized point cloud created by GAN. Note that during whole training and test process, we did not give network any information about object parts, but generator tends to learn different colour patterns for different parts by itself.