Local Field Scaling for Point Surface Normal Estimation

In this paper, we propose a refined technique for point surface normal estimation utilizing local receptive field scaling. Traditional methods often struggle with artifact patterns like gradients and stripes, compromising accuracy. To address this issue, the proposed method integrates relative positioning for streamlined weighted surface approximation and local farthest point distance normalization to address artifact density and sparsity challenges. In addition, our approach enhances network training efficiency. Experimental results on public datasets confirm that our method outperforms existing techniques in the sense of both accuracy and robustness.