Abstract Development of LiDAR and depth camera makes it easily to extract the point cloud data of practical items. However, some drawbacks, such as sparsity or loss of details of the point cloud, are evident. Therefore, quite different from the methods as developed so far which usually reconstructed incomplete point cloud either in terms of GAN-based or autoencoder-based networks, respectively. In this paper, we propose PointALCR, which combines GAN-based and autoencoder-based frameworks with contrastive regularization in order to improve the representative and generative abilities for completion of the point cloud. A module named Adversarial Latent GAN be employed for learning a latent space of input/target point cloud representation and extending the generative and discriminative abilities on GAN training procedures. Contrastive regularization ensures that the reconstructed items to be close to the ground truth and far from the incomplete input in feature space. Experimental results demonstrate that PointALCR has the capabilities better than previous methods on challenging point cloud completion tasks.
展开▼
