Enhancing Local Feature Learning for 3D Point Cloud Processing using Unary-Pairwise Attention

Abstract

We present a simple but effective attention named the unary-pairwise attention (UPA) for modeling the relationship between 3D point clouds. Our idea is motivated by the analysis that the standard self-attention (SA) that operates globally tends to produce almost the same attention maps for different query positions, revealing difficulties for learning query-independent and query-dependent information jointly. Therefore, we reformulate the SA and propose query-independent (Unary) and query-dependent (Pairwise) components to facilitate the learning of both terms. In contrast to the SA, the UPA ensures query dependence via operating locally. Extensive experiments show that the UPA outperforms the SA consistently on various point cloud understanding tasks including shape classification, part segmentation, and scene segmentation. Moreover, simply equipping the popular PointNet++ method with the UPA even outperforms or is on par with the stateof-the-art attention-based approaches. In addition, the UPA systematically boosts the performance of both standard and modern networks when it is integrated into them as a compositional module.

Enhancing Local Geometry Learning for 3D Point Cloud via Decoupling Convolution

“…On the other hand, motivated by the success of self-attention (Vaswani et al, 2017), attention mechanisms (Vaswani et al, 2017) are widely adopted to dynamically compute connectivity using the point feature similarity. For instance, some works adopt the dot product for similarity measure (Yang et al, 2019;Yan et al, 2020) whereas edges are used in other works (Wang et al, 2019a;Zhao et al, 2021;Xiu et al, 2022).…”

“…Some works (Hua et al, 2018;Atzmon et al, 2018;Thomas et al, 2019;Mao et al, 2019) explicitly introduce regular convolutional kernels to which the points are projected while the others dynamically predict the convolution filter using various features (Wang et al, 2018;Liu et al, 2019;Li et al, 2018b;Wu et al, 2019;Wang et al, 2019b;Simonovsky and Komodakis, 2017;Li et al, 2019;Liu et al, 2020;Xu et al, 2021;Xiang et al, 2021). Recently, inspired by the success of self-attention (Vaswani et al, 2017), a line of research incorporates the attention mechanism into networks (Wang et al, 2019a;Zhao et al, 2019Zhao et al, , 2021Xiu et al, 2022).…”

Enhancing Local Feature Learning Using Diffusion for 3D Point Cloud Understanding

“…In [43], such a similarity is used for describing connectivity among neighboring points; consequently, the information exchange is guided by edge information. Such an operation is often coupled with the attention mechanism [55,56,49], which converts edge into normalized spatial weights. However, the forced conversion may lose rich structural information (e.g., smoothness) contained in the edges.…”