Point Cloud Upsampling by Generative Adversarial Network with Skip-attention

Point clouds are a very popular way of representing 3D data. The development of various advanced devices has made it easily accessible. However, the acquired point cloud data usually has the following characteristics: the number of points is not easily controlled, sparse and non-uniform. These characteristics make it difficult to apply the acquired point cloud data directly to various tasks. To effectively address these issues, we propose a new method based on generative adversarial networks to implement upsampling pre-processing of point clouds. It is possible to easily upsample the number of points in the point cloud to our desired value and the obtained point cloud data can be very homogeneous while maintaining the original contours. In detail, we have introduced Skip-attention to our generator, which allows the network to effectively fuse the local and global features of the point cloud, and in addition to this, we have used PointNet-Mix as our discriminator, a simple and lightweight structure that works well with our generator. Extensive qualitative and quantitative experiments have demonstrated that the upsampling data obtained using our method can achieve equally competitive results.