[Paper Review] VA-GCN: A Vector Attention Graph Convolution Network for learning on Point Clouds

Haotian Hu, Fanyi Wang+1 more|arXiv (Cornell University)|Jun 1, 2021

3D Shape Modeling and Analysis35 references4 citations

TL;DR

This paper proposes VA-GCN, a novel graph convolutional network for point cloud learning that introduces a Vector Attention Convolution (VAConv) module to explicitly model local geometric relationships using elevation and azimuth angles between points. By fusing high-dimensional edge features with low-dimensional geometric attention and stacking multi-scale VAConv modules, VA-GCN achieves state-of-the-art performance, improving accuracy by 2.4% on ModelNet40 and setting new SOTA on S3DIS and ShapeNet benchmarks.

ABSTRACT

Owing to the development of research on local aggregation operators, dramatic breakthrough has been made in point cloud analysis models. However, existing local aggregation operators in the current literature fail to attach decent importance to the local information of the point cloud, which limits the power of the models. To fit this gap, we propose an efficient Vector Attention Convolution module (VAConv), which utilizes K-Nearest Neighbor (KNN) to extract the neighbor points of each input point, and then uses the elevation and azimuth relationship of the vectors between the center point and its neighbors to construct an attention weight matrix for edge features. Afterwards, the VAConv adopts a dual-channel structure to fuse weighted edge features and global features. To verify the efficiency of the VAConv, we connect the VAConvs with different receptive fields in parallel to obtain a Multi-scale graph convolutional network, VA-GCN. The proposed VA-GCN achieves state-of-the-art performance on standard benchmarks including ModelNet40, S3DIS and ShapeNet. Remarkably, on the ModelNet40 dataset for 3D classification, VA-GCN increased by 2.4% compared to the baseline.

Motivation & Objective

• To address the limitation of existing local aggregation operators in point cloud models, which fail to adequately capture and weight local geometric information.
• To develop a mechanism that explicitly models the directional and angular relationships between a point and its neighbors to improve feature representation.
• To design a multi-scale, dual-channel architecture that fuses high-dimensional edge features with low-dimensional geometric structure for richer feature learning.
• To achieve state-of-the-art performance on standard 3D point cloud benchmarks including ModelNet40, S3DIS, and ShapeNet for classification, segmentation, and part segmentation tasks.

Proposed method

• VAConv uses K-Nearest Neighbor (KNN) to identify neighbor points for each input point, forming a directed graph for local neighborhood representation.
• It computes elevation and azimuth angles from relative vectors between the center point and its neighbors to construct a learnable attention weight matrix for edge features.
• The module employs a dual-channel structure: one branch processes weighted edge features using VAConv, and the other fuses global features with relative geometric information.
• The VA-GCN architecture stacks VAConv modules with varying receptive fields in parallel to capture multi-scale local structures.
• The network integrates EdgeConv and VAConv in a dual-pathway design, enabling joint learning of geometric topology and semantic edge features.
• Multi-Sample Inference (MSI) is applied during testing to further boost performance on ModelNet40.

Experimental results

Research questions

• RQ1 How can local geometric relationships in point clouds be explicitly modeled to improve feature representation in deep learning models?
• RQ2 What is the impact of incorporating directional vector information (elevation and azimuth) into attention mechanisms for edge feature weighting?
• RQ3 Can a dual-channel architecture that fuses high-dimensional edge features and low-dimensional geometric structure improve performance on point cloud tasks?
• RQ4 How does multi-scale aggregation using parallel VAConv modules with different receptive fields affect model accuracy and generalization?
• RQ5 To what extent does Multi-Sample Inference (MSI) enhance performance on point cloud classification benchmarks?

Key findings

• VA-GCN achieves a 93.5% overall accuracy on ModelNet40, representing a 2.4% improvement over the baseline and setting a new state-of-the-art score.
• On the S3DIS dataset for indoor semantic segmentation, VA-GCN achieves a 56.9% mIoU, surpassing the previous best result by 0.8%.
• For 3D part segmentation on ShapeNet, VA-GCN achieves 85.5% instance mIoU and 82.6% class mIoU, outperforming DGCNN by 0.3%.
• Abllation studies confirm that the dual-channel structure improves accuracy by 0.7% over single-channel variants, and the parallel multi-scale VAConv design boosts performance by 0.9% compared to single-branch configurations.
• The proposed Multi-Sample Inference (MSI) strategy further enhances the ModelNet40 result, achieving the highest reported score to date.