Abstract:Point cloud upsampling focuses on generating a dense, uniform and proximity-to-surface point set. Most previous approaches accomplish these objectives by carefully designing a single-stage network, which makes it still challenging to generate a high-fidelity point distribution. Instead, upsampling point cloud in a coarse-to-fine manner is a decent solution. However, existing coarse-to-fine upsampling methods require extra training strategies, which are complicated and time-consuming during the training. In this paper, we propose a simple yet effective cascaded refinement network, consisting of three generation stages that have the same network architecture but achieve different objectives. Specifically, the first two upsampling stages generate the dense but coarse points progressively, while the last refinement stage further adjust the coarse points to a better position. To mitigate the learning conflicts between multiple stages and decrease the difficulty of regressing new points, we encourage each stage to predict the point offsets with respect to the input shape. In this manner, the proposed cascaded refinement network can be easily optimized without extra learning strategies. Moreover, we design a transformer-based feature extraction module to learn the informative global and local shape context. In inference phase, we can dynamically adjust the model efficiency and effectiveness, depending on the available computational resources. Extensive experiments on both synthetic and real-scanned datasets demonstrate that the proposed approach outperforms the existing state-of-the-art methods.

HFCI-PU: High-Frequency Compensation Interpolation Network for Point Cloud Upsampling

Point cloud upsampling via a coarse-to-fine network with transformer-encoder

PC2-PU: Patch Correlation and Point Correlation for Effective Point Cloud Upsampling

PU-Net: Point Cloud Upsampling Network

PC$^2$-PU: Patch Correlation and Point Correlation for Effective Point Cloud Upsampling

PU-CTG: A Point Cloud Upsampling Network Using Transformer Fusion and GRU Correction

Point Cloud Upsampling via Cascaded Refinement Network

Point Cloud Upsampling Via Disentangled Refinement

A Two-Stage Transformer-Based Point Cloud Upsampling

PU-Transformer: Point Cloud Upsampling Transformer

PU-Flow: a Point Cloud Upsampling Network with Normalizing Flows

Upu-Snet: Siamese Network for Unsupervised Point Cloud Upsampling Based on Spatial-Aware Transformers

SSPU-Net: A Structure Sensitive Point Cloud Upsampling Network with Multi-Scale Spatial Refinement

HFF-Net: Hierarchical Feature Fusion Network for Point Cloud Generation with Point Transformers

Data-driven Upsampling of Point Clouds

CPP-Net: Embracing Multi-Scale Feature Fusion into Deep Unfolding CP-PPA Network for Compressive Sensing

Neural Points: Point Cloud Representation with Neural Fields for Arbitrary Upsampling

RepKPU: Point Cloud Upsampling with Kernel Point Representation and Deformation

VPU: A Video-Based Point Cloud Upsampling Framework

SmoothNet: Smooth Point Cloud Up-sampling

Invertible Residual Neural Networks with Conditional Injector and Interpolator for Point Cloud Upsampling