由於實務上在進行點雲部件切割時無法保證輸入點雲始終能夠保持同一方向，因此使模型具有足夠的泛化性並具有旋轉強健性至關重要，本論文提出一種全新的旋轉強健性技術──神經正切逼近法，我們利用神經正切核來尋找旋轉角度加入訓練資料，它能夠避免對於複雜的數學與幾何學知識的要求，同時可以顯著的降低計算成本以及記憶體空間。在ShapeNetPart上進行的實驗表明了神經正切逼近法可以在保留對未旋轉點雲的高準確率的同時使模型具有旋轉強健性的能力，與ART-Point相比，我們在搜尋旋轉角度的速度上快了將近9倍，且記憶體的使用量也減少了將近1倍，同時實驗也表明了神經正切逼近法對於未旋轉點雲及旋轉後點雲的準確率皆高於ART-Point，這也顯示了我們所提出的神經正切逼近法與其他state-of-the-art方法具有可比較性。;In practical point cloud part segmentation applications, it is often impossible to ensure that the input point clouds keep the same orientation. Therefore, it is crucial for the model to generalize well and possess rotation robustness. This thesis proposes a novel method for achieving rotation robustness: the Neural Tangent Approximation Method. By utilizing the neural tangent kernel, we integrate augmented data with rotation angles into the training data. This approach avoids the need for complex mathematical and geometrical knowledge, significantly reducing computational costs and memory usage. Experiments conducted on the ShapeNetPart dataset demonstrate that the Neural Tangent Approximation Method maintains high accuracy for non-rotated point clouds while enhancing robustness for rotated inputs. Compared to ART-Point, our method is nearly nine times faster at searching for rotation angles and uses about half as much memory. Furthermore, our experiments show that our method surpasses ART-Point in accuracy for both non-rotated and rotated point clouds, achieving state-of-the-art performance.
