Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds-CFANZ编程社区

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds

Yunhao Cui, Yi An, Member, IEEE, Wei Sun, Huosheng Hu, Senior Member, IEEE, and Xueguan Song

年份：2020

期刊：IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT

IF：3.658

1、创新

1、将深度可分离卷积引入到PointNet等网络中，实现了轻量级+高效，Lightweight Module 。

2、为了弥补深度可分离卷积导致的精度下降的问题，引入了channel attention模块，提高了精度，Attention Module。

3、将两个模块（LM+AM）结合起来，使轻量级和高精度之间达到了平衡。

深度可分离卷积可以参考：这篇

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_3D

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_3D_02

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_3D_03

常规卷积的参数量：

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_3D_04

深度可分离卷积的参数量：

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_卷积_05

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_3D_06

M代表每一个通道上特征的均值：

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_权重_07

D代表每一个通道上特征的分布标准差：

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_3D_08

每一个通道上的"权重"：

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_3D_09

最终得到注意力权重：

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_权重_10

根据权重更新通道的特征：

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_卷积_11

借助于PointNet的结构，将其中的mlp替换为本文提出的LAM：

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_卷积_12

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_3D_13

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_权重_14

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_权重_15

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_权重_16

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_卷积_17

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point Clouds_卷积_18

0 条评论