GT-MilliNoise: Graph transformer for point-wise denoising of indoor millimetre-wave point clouds

Abstract

Millimetre-wave (mmWave) radars are gaining popularity thanks to their low cost and robustness in low-visibility conditions. However, the 3D point clouds they produce are sparser and noisier than those from LiDARs and depth cameras. These differences create challenges when applying existing methods, originally designed for dense point clouds, to mmWave data. Specifically, there is a gap in point-level precision tasks, such as full point cloud denoising for mmWave data, partly due to the lack of fully annotated datasets. In this work, we employ the MilliNoise dataset, a fully annotated indoor mmWave point clouds dataset, to advance the understanding of mmWave point clouds denoising via two main steps: (i) we carry out an experimental analysis of the most common point cloud processing approaches and show their limitations in exploring the local-to-global structures in sparse and noisy point clouds; (ii) in light of the identified limitations, we propose a graph-based transformer architecture, denoted as GT-MilliNoise, composed of two main blocks to effectively leverage both the temporal and geometric structures of the data: a Temporal block leverages the sparsity of data to learn the dynamic behaviour of the points; a Geometric block, uses a point-wise attention mechanism to form representative neighbourhoods for feature extraction. The experimental results obtained in the MilliNoise dataset show that our proposed GT-MilliNoise architecture outperforms the state-of-the-art both qualitatively and quantitatively. Specifically, it achieves 75% accuracy (5% gain compared to the state-of-the-art), and a significantly low Earth Mover’s distance value of 0.193.