Reconstruction-Aware Prior Distillation for Semi-supervised Point Cloud Completion

by   Zhaoxin Fan, et al.
Renmin University of China
Tsinghua University

Point clouds scanned by real-world sensors are always incomplete, irregular, and noisy, making the point cloud completion task become increasingly more important. Though many point cloud completion methods have been proposed, most of them require a large number of paired complete-incomplete point clouds for training, which is labor exhausted. In contrast, this paper proposes a novel Reconstruction-Aware Prior Distillation semi-supervised point cloud completion method named RaPD, which takes advantage of a two-stage training scheme to reduce the dependence on a large-scale paired dataset. In training stage 1, the so-called deep semantic prior is learned from both unpaired complete and unpaired incomplete point clouds using a reconstruction-aware pretraining process. While in training stage 2, we introduce a semi-supervised prior distillation process, where an encoder-decoder-based completion network is trained by distilling the prior into the network utilizing only a small number of paired training samples. A self-supervised completion module is further introduced, excavating the value of a large number of unpaired incomplete point clouds, leading to an increase in the network's performance. Extensive experiments on several widely used datasets demonstrate that RaPD, the first semi-supervised point cloud completion method, achieves superior performance to previous methods on both homologous and heterologous scenarios.


page 1

page 2

page 3

page 4


Leveraging Single-View Images for Unsupervised 3D Point Cloud Completion

Point clouds captured by scanning devices are often incomplete due to oc...

Are All Point Clouds Suitable for Completion? Weakly Supervised Quality Evaluation Network for Point Cloud Completion

In the practical application of point cloud completion tasks, real data ...

CP3: Unifying Point Cloud Completion by Pretrain-Prompt-Predict Paradigm

Point cloud completion aims to predict complete shape from its partial o...

Hyperspherical Embedding for Point Cloud Completion

Most real-world 3D measurements from depth sensors are incomplete, and t...

Circular Accessible Depth: A Robust Traversability Representation for UGV Navigation

In this paper, we present the Circular Accessible Depth (CAD), a robust ...

Completing Partial Point Clouds with Outliers by Collaborative Completion and Segmentation

Most existing point cloud completion methods are only applicable to part...

Point-Cloud Completion with Pretrained Text-to-image Diffusion Models

Point-cloud data collected in real-world applications are often incomple...

Please sign up or login with your details

Forgot password? Click here to reset