FatNet: A Feature-attentive Network for 3D Point Cloud Processing

Chaitanya Kaul; Nicholas Edwin Pears; Suresh Manandhar

FatNet: A Feature-attentive Network for 3D Point Cloud Processing

Chaitanya Kaul, Nicholas Edwin Pears, Suresh Manandhar

Computer Science

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

Abstract

The application of deep learning to 3D point clouds is challenging due to its lack of order. Inspired by the point embeddings of PointNet and the edge embeddings of DGCNNs, we propose three improvements to the task of point cloud analysis. First, we introduce a novel feature-attentive neural network layer, a FAT layer, that combines both global point-based features and local edge-based features in order to generate better embeddings. Second, we find that applying the same attention mechanism across two different forms of feature map aggregation, max pooling and average pooling, gives better performance than either alone. Third, we observe that residual feature reuse in this setting propagates information more effectively between the layers, and makes the network easier to train. Our architecture achieves state-of-the-art results on the task of point cloud classification, as demonstrated on the ModelNet40 dataset, and an extremely competitive performance on the ShapeNet part segmentation challenge.

Original language	English
Title of host publication	25th International Conference on Pattern Recognition
Publisher	Springer
Publication status	Published - 10 Jan 2021
Event	25th International Conference on Pattern Recognition - Milan, Italy Duration: 10 Jan 2021 → 15 Jan 2021 https://www.micc.unifi.it/icpr2020/

Conference

Conference	25th International Conference on Pattern Recognition
Abbreviated title	ICPR 2020
Country/Territory	Italy
City	Milan
Period	10/01/21 → 15/01/21
Internet address	https://www.micc.unifi.it/icpr2020/

Keywords

3D Point Cloud, deep network, 3D shape classification, 3D shape retrieval, 3D shape segmentation

Cite this

@inproceedings{94db77733d1a457180565c42fbfa6301,

title = "FatNet: A Feature-attentive Network for 3D Point Cloud Processing",

abstract = "The application of deep learning to 3D point clouds is challenging due to its lack of order. Inspired by the point embeddings of PointNet and the edge embeddings of DGCNNs, we propose three improvements to the task of point cloud analysis. First, we introduce a novel feature-attentive neural network layer, a FAT layer, that combines both global point-based features and local edge-based features in order to generate better embeddings. Second, we find that applying the same attention mechanism across two different forms of feature map aggregation, max pooling and average pooling, gives better performance than either alone. Third, we observe that residual feature reuse in this setting propagates information more effectively between the layers, and makes the network easier to train. Our architecture achieves state-of-the-art results on the task of point cloud classification, as demonstrated on the ModelNet40 dataset, and an extremely competitive performance on the ShapeNet part segmentation challenge.",

keywords = "3D Point Cloud, deep network, 3D shape classification, 3D shape retrieval, 3D shape segmentation",

author = "Chaitanya Kaul and Pears, {Nicholas Edwin} and Suresh Manandhar",

year = "2021",

month = jan,

day = "10",

language = "English",

booktitle = "25th International Conference on Pattern Recognition",

publisher = "Springer",

address = "Germany",

note = "25th International Conference on Pattern Recognition, ICPR 2020 ; Conference date: 10-01-2021 Through 15-01-2021",

url = "https://www.micc.unifi.it/icpr2020/",

}

TY - GEN

T1 - FatNet: A Feature-attentive Network for 3D Point Cloud Processing

AU - Kaul, Chaitanya

AU - Pears, Nicholas Edwin

AU - Manandhar, Suresh

PY - 2021/1/10

Y1 - 2021/1/10

N2 - The application of deep learning to 3D point clouds is challenging due to its lack of order. Inspired by the point embeddings of PointNet and the edge embeddings of DGCNNs, we propose three improvements to the task of point cloud analysis. First, we introduce a novel feature-attentive neural network layer, a FAT layer, that combines both global point-based features and local edge-based features in order to generate better embeddings. Second, we find that applying the same attention mechanism across two different forms of feature map aggregation, max pooling and average pooling, gives better performance than either alone. Third, we observe that residual feature reuse in this setting propagates information more effectively between the layers, and makes the network easier to train. Our architecture achieves state-of-the-art results on the task of point cloud classification, as demonstrated on the ModelNet40 dataset, and an extremely competitive performance on the ShapeNet part segmentation challenge.

AB - The application of deep learning to 3D point clouds is challenging due to its lack of order. Inspired by the point embeddings of PointNet and the edge embeddings of DGCNNs, we propose three improvements to the task of point cloud analysis. First, we introduce a novel feature-attentive neural network layer, a FAT layer, that combines both global point-based features and local edge-based features in order to generate better embeddings. Second, we find that applying the same attention mechanism across two different forms of feature map aggregation, max pooling and average pooling, gives better performance than either alone. Third, we observe that residual feature reuse in this setting propagates information more effectively between the layers, and makes the network easier to train. Our architecture achieves state-of-the-art results on the task of point cloud classification, as demonstrated on the ModelNet40 dataset, and an extremely competitive performance on the ShapeNet part segmentation challenge.

KW - 3D Point Cloud, deep network, 3D shape classification, 3D shape retrieval, 3D shape segmentation

M3 - Conference contribution

BT - 25th International Conference on Pattern Recognition

PB - Springer

T2 - 25th International Conference on Pattern Recognition

Y2 - 10 January 2021 through 15 January 2021

ER -