SQN
type
status
date
slug
summary
tags
category
icon
password
论文SQN: Weakly-Supervised Semantic Segmentation of Large-Scale 3D Point Clouds的阅读笔记
📝 SQN
Overview
两个主要组件:
- a point local feature extractor to learn diverse visual patterns;
- a flexible point feature query network to collect as many as possible relevant semantic features for weakly-supervised training
Point Local Feature Extractor
- 可替换为任意的backbone
- 降采样操作提取多尺度局部特征
Point Feature Query Network
- 将有label的点与其周围点的特征聚合,从而有label的点的训练信号可以分析给它周围的点
Searching Spatial Neighbouring Point Features
search the nearest K points in each of the previous 4-level encoded features
Interpolating Query Point Features
apply the trilinear interpolation method to compute a feature vector for p
Inferring Query Point Semantics
feed the unique and representative feature vector into a series of MLPs
🤗 Summary
基于本论文实现了SQN pytorch-lightning的版本.
SQN_pl
BigCiLeng • Updated Jun 12, 2023
📎 参考文章
- 一些引用
- 引用文章