Answers are grounded in this paper and cite their sources.
Real-scanned raw point clouds face dual challenges: sparse non-uniform distribution and significant noise/outliers. While many upsampling methods address sparsity, they remain highly noise-sensitive. That is, they fail to solve both issues. To fill this gap, we propose Robust-APU, an end-to-end unified framework integrating arbitrary upsampling and denoising. Our approach synergistically balances the two tasks: first, it interpolates the sparse point cloud to lay the upsampling foundation, then uses a dedicated noise-resistant module for preliminary denoising. Next, we design a weight-adaptive upsampling module to extract upsampled features from the original input. Though these features contain noise, they encode rich geometric information. Finally, a position refinement module fuses features from the denoising module and the upsampling module to enhance upsampled point accuracy and eliminates residual noise, boosting both upsampling quality and noise robustness. Extensive experiments show that Robust-APU outperforms state-of-the-arts at various upsampling rates. Specifically, on the PU-GAN dataset under the Gaussian noise level of 2%, it achieves 34.6% and 39.2% improvement in P2F metric, proving its ability to deliver high-quality upsampling while suppressing noise effectively.
No takes yet. Share an insight, caveat, or question.
Yu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: