[Paper Review] GAPNet: Graph Attention based Point Neural Network for Exploiting Local Feature of Point Cloud

Can Chen, Luca Zanotti Fragonara+1 more|arXiv (Cornell University)|May 21, 2019

3D Shape Modeling and Analysis40 citations

TL;DR

GAPNet introduces multi-head graph attention on point clouds to learn local geometric features, achieving state-of-the-art results on ModelNet40 classification and ShapeNet part segmentation. It integrates GAPLayer with attention pooling into an MLP framework.

ABSTRACT

Exploiting fine-grained semantic features on point cloud is still challenging due to its irregular and sparse structure in a non-Euclidean space. Among existing studies, PointNet provides an efficient and promising approach to learn shape features directly on unordered 3D point cloud and has achieved competitive performance. However, local feature that is helpful towards better contextual learning is not considered. Meanwhile, attention mechanism shows efficiency in capturing node representation on graph-based data by attending over neighboring nodes. In this paper, we propose a novel neural network for point cloud, dubbed GAPNet, to learn local geometric representations by embedding graph attention mechanism within stacked Multi-Layer-Perceptron (MLP) layers. Firstly, we introduce a GAPLayer to learn attention features for each point by highlighting different attention weights on neighborhood. Secondly, in order to exploit sufficient features, a multi-head mechanism is employed to allow GAPLayer to aggregate different features from independent heads. Thirdly, we propose an attention pooling layer over neighbors to capture local signature aimed at enhancing network robustness. Finally, GAPNet applies stacked MLP layers to attention features and local signature to fully extract local geometric structures. The proposed GAPNet architecture is tested on the ModelNet40 and ShapeNet part datasets, and achieves state-of-the-art performance in both shape classification and part segmentation tasks.

Motivation & Objective

• Motivate and address the challenge of leveraging fine-grained local features in irregular point clouds.
• Propose GAPNet with a GAPLayer that uses self-attention and neighbor-attention over k-NN graphs.
• Introduce an attention pooling layer to obtain local signature and improve robustness.
• Integrate GAPLayer and pooling into stacked MLPs for improved local geometric feature extraction.
• Demonstrate state-of-the-art performance on ModelNet40 for shape classification and ShapeNet part segmentation.

Proposed method

• Construct a directed k-NN graph to model local structure for each point.
• Compute self- and local-attention coefficients via a single-head GAPLayer using a LeakyReLU fusion and softmax normalization.
• Aggregate attended neighbor features to form attention features.
• Extend to multi-head GAPLayer by concatenating outputs from M heads.
• Apply an attention pooling layer over neighbor features to obtain a local signature.
• Integrate GAPLayer and pooling into an architecture with shared MLPs for classification and segmentation tasks.

Experimental results

Research questions

• RQ1 Can graph attention over local neighborhoods improve local feature learning in point clouds for classification and segmentation?
• RQ2 Does multi-head GAPLayer better capture diverse local representations than a single attention head?
• RQ3 What is the impact of the attention pooling layer on robustness and performance?
• RQ4 How does GAPNet compare to state-of-the-art on ModelNet40 and ShapeNet Part datasets?

Key findings

• GAPNet achieves state-of-the-art performance on ModelNet40 shape classification, outperforming prior methods including DGCNN by 0.2% accuracy in the reported setup.
• On ShapeNet Part, GAPNet yields competitive mean IoU and part-wise accuracy, with strong performance across multiple categories.
• An ablation study shows that the attention pooling layer adds about 0.6% accuracy, and the GAPLayer’s learned attention (vs constant coefficients) contributes about 0.7% accuracy.
• Varying the number of heads and encoding channels reveals a balance where 4 heads with 16 channels provides strong performance; too many heads/channels can degrade results.