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

Point cloud upsampling (PCU) aims to transform sparse and unevenly distributed point clouds into dense and uniform counterparts with intricate geometric details of real-world objects. Existing networks fall short in compensating for the low-pass nature of transformer and fail to establish robust global context correlations and local channel correlations without highlighting stage specificity. To address these challenges, we introduce the high-frequency compensation interpolation network for PCU (HFCI-PU), which leverages patch-based upsampling on the target object surfaces. Specifically, we design a novel interpolation strategy that compensates for the high-frequency points by interpolating them in a multistage manner to address the high-frequency point decay in transformers, and stage specificity is highlighted from the perspective of task-driven to minimize redundancy. In addition, we design a multiscale dense feature extraction (MDFE) to build global correlations between features during the leaning process and a channel correlated point transformer (CCPT) to construct local channel correlations. Qualitative and quantitative evaluations on four datasets demonstrate the competitiveness of our approach, showcasing superior reconstruction quality and proximity to the target object surface. Our code will be released in the public repository at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/xiaolongTang163/HFCI-PU</uri>.