一種應用於三維點雲資料的三維幾何特徵提取方法

Abstract

近年來在深度學習領域，許多卷積運算子的提出在三維幾何資料中有著很好的表現。有些運算子只拿三維座標作為輸入便能有很好的表現。然而，旋轉不變性對於三維資料來說是個有挑戰性的課題。有些運算子的表現在經過旋轉之後往往會變差。在這篇論文中，我們提出了一個架構，能根據點集合的中心和參考點抽取特定不受旋轉影響的幾何特徵，且同時搜尋不同數量的最近鄰近來組成不同的點集合。在我們的設計裡，我們在特定的點雲資料中，同時搜尋不同數量的最近鄰居，然後逐步減少點雲的取樣好以取得較多不受旋轉影響的特徵資訊。我們實驗在不同且知名的點雲資料集，例如：物件辨識、物件分割等等。我們還針對我們的模型結構去做分析，而結果說明我們的設計還有所抽取的特徵著實有改善模型的表現。實驗也證實我們的方法即便在經過旋轉過後的資料依然能生成一樣的結果，且能有最先進的表現。其在ModelNet40 這個物件辨識資料集上能達到平均90.4% 的預測正確率，而在含有背景資訊的ScanObjectNN 這個物件辨識的資料集上也能達到平均76.2% 的預測正確率，在ShapeNet 這個物件分割的資料集上則可達到79.2% 的mIoU。

Recently, a lot of convolution operators have a good performance when dealing with 3D geometric data. Some of them simply take the point set located in 3D coordinates as inputs and are able to achieve a good performance. However, rotation-invariance is a challenging task for point cloud data. Analyses for some operators are prone to degrade after arbitrary rotation. In the thesis, we propose a novel framework that extracts rotation-invariant features relative to the centroid and the reference point in a local point set, and takes different numbers of nearest neighbors simultaneously to formulate distinct local point sets and centroids. We experiment with different well-known point cloud tasks, such as object classification as well as part segmentation. We also conduct design analysis on our model architecture, and the results show that our design indeed improves the performance. Experiments demonstrate that our method generates consistent results on randomly rotated data and achieves state-of-the-art performances without any rotation augmentation. Our model is able to reach 90.4% accuracy on average on the ModelNet40 dataset. On the ScanObjectNN dataset, which includes information of background, our model is able to reach 76.2% accuracy on average. On the ShapeNet dataset, which is a part segmentation task, our model is able to achieve 79.2% mIOU.