type
status
date
slug
summary
tags
category
icon
password
Last edited time
Jun 16, 2024 07:19 PM
点云预处理论文素材
📝 主旨内容
贡献(创新点)
- 提出了一种针对MVTec 3D-AD数据集的用于工业异常检测的多模态预处理方法
- 我们的方法结合最先进的异常检测方法取得了新的sota成绩
- 与传统预处理方法相比处理时间下降了151倍,CPU占用下降了5倍,内存消耗下降了18倍
- 我们的方法能够很好地融合现有的异常检测方法,形成一个端到端的具备实时性能的多模态工业异常检测框架,与传统异常检测方法相比FPS最高提升120倍
比较的方法
方法名 | 发表会议/期刊 | 时间 | ㅤ | 预处理方法 |
3D-ADS | CVPRW | 2023 | BTF | |
CPMF | Pattern Recognition(PR) | 2023 | BTF | |
AST | WACV | 2023 | AST | |
M3DM | CVPR | 2023 | BTF | |
EasyNet | MM | 2023 | BTF | |
Shape-Guild | ICML | 2023 | BTF | |
3DSR | WACV | 2024 | 3DSR |
配图
不可能铁三角
具体提升/下降的倍数
随CPU性能的变化(附录)
预处理结果对比(更好的可视化方式?)
ㅤ | 类1 | 类2 | 类3 | 类4 | 类5 | 类6 |
RGB | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
Point Cloud | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
去除平面 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
padded | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
去除多余联通块 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ours | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
异常检测结果可视化
流程可视化+算法实现
测试强行端到端的FPS(各种方法仅更换预处理部分)
探索阈值设置对时间 功耗 效果的影响
表格
附件
更多可视化结果
预处理结果展示
异常检测结果展示
全部表格数据
CPU 内存占用表格
时间
性能
参考文献
数据集
MVtec 3dad
MVtec ad
Eyecandies
三维异常检测
3D-ADS
M3DM
CPMF
EasyNet
Shaped_Guilded
二维异常检测
A Unified Model for Multi-class Anomaly Detection
patchcore
simplenet
异常检测指标
AUPRO
BTF原理解析
RANSAC
DB-Scan
open3d
2D异常检测?
分析原因
BTF
预处理有效,得看具体方法
BTF预处理日志
M3DM训练并测试双库版本并保存该功能以进行 UFF 训练
Train and test the double lib version and save the feature for UFF training:
Train the UFF:
Train and test the full setting with the following command:
V100上执行提供的预训练模型日志
V100上执行提供的预训练模型日志(无预处理)
预处理大幅减少基于Memery Bank方法的内存占用
M3DM I-AUROC
ㅤ | Bagel | Cable_Gland | Carrot | Cookie | Dowel | Foam | Peach | Potato | Rope | Tire | Mean |
双库版本 | 0.996 | 0.829 | 0.982 | 0.946 | 0.899 | 0.865 | 0.877 | 0.944 | 0.957 | 0.794 | 0.909 |
双库版本(processed) | 0.994 | 0.846 | 0.981 | 0.968 | 0.901 | 0.869 | 0.930 | 0.951 | 0.968 | 0.848 | 0.9256 |
双库版本(raw) | 0.989 | 0.884 | 0.980 | 0.968 | 0.892 | 0.841 | 0.943 | 0.946 | 0.993 | 0.836 | 0.9272 |
双库版本(Ours) | 0.984 | 0.802 | 0.981 | 0.941 | 0.936 | 0.852 | 0.938 | 0.945 | 0.94 | 0.963 | 0.918 |
公开预训练权重的融合版本 | 0.998 | 0.901 | 0.972 | 0.972 | 0.926 | 0.897 | 0.958 | 0.906 | 0.976 | 0.811 | 0.932 |
论文报告 | 0.994 | 0.909 | 0.972 | 0.976 | 0.960 | 0.942 | 0.973 | 0.899 | 0.972 | 0.850 | 0.945 |
同环境复现 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
公开预训练权重的融合版本(不预处理) | 0.999 | 0.946 | 0.976 | 0.955 | 0.899 | 0.869 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
M3DM P-AUROC
ㅤ | Bagel | Cable_Gland | Carrot | Cookie | Dowel | Foam | Peach | Potato | Rope | Tire | Mean |
双库版本 | 0.994 | 0.992 | 0.996 | 0.978 | 0.981 | 0.982 | 0.996 | 0.994 | 0.996 | 0.995 | 0.99 |
双库版本(processed) | 0.994 | 0.992 | 0.996 | 0.978 | 0.981 | 0.982 | 0.996 | 0.994 | 0.997 | 0.995 | 0.9905 |
双库版本(raw) | 0.993 | 0.980 | 0.996 | 0.967 | 0.984 | 0.983 | 0.995 | 0.996 | 0.993 | 0.990 | 0.9877 |
双库版本(Ours) | 0.993 | 0.992 | 0.996 | 0.978 | 0.981 | 0.982 | 0.996 | 0.994 | 0.997 | 0.995 | 0.99 |
公开预训练权重的融合版本 | 0.994 | 0.993 | 0.997 | 0.985 | 0.985 | 0.98 | 0.996 | 0.994 | 0.997 | 0.995 | 0.992 |
论文报告 | 0.995 | 0.993 | 0.997 | 0.985 | 0.985 | 0.984 | 0.996 | 0.994 | 0.997 | 0.996 | 0.992 |
同环境复现 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
公开预训练权重的融合版本(不预处理) | 0.995 | 0.988 | 0.997 | 0.978 | 0.992 | 0.988 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
M3DM AUPRO
ㅤ | Bagel | Cable_Gland | Carrot | Cookie | Dowel | Foam | Peach | Potato | Rope | Tire | Mean |
双库版本 | 0.996 | 0.968 | 0.978 | 0.935 | 0.93 | 0.927 | 0.977 | 0.966 | 0.969 | 0.972 | 0.959 |
双库版本(processed) | 0.968 | 0.969 | 0.978 | 0.935 | 0.928 | 0.926 | 0.976 | 0.966 | 0.970 | 0.971 | 0.9587 |
双库版本(raw) | 0.971 | 0.931 | 0.973 | 0.930 | 0.923 | 0.922 | 0.974 | 0.973 | 0.948 | 0.954 | 0.9499 |
双库版本(Ours) | 0.966 | 0.968 | 0.978 | 0.932 | 0.931 | 0.927 | 0.976 | 0.967 | 0.97 | 0.972 | 0.959 |
公开预训练权重的融合版本 | 0.969 | 0.97 | 0.978 | 0.948 | 0.94 | 0.918 | 0.976 | 0.965 | 0.971 | 0.973 | 0.961 |
论文报告 | 0.970 | 0.971 | 0.979 | 0.950 | 0.941 | 0.932 | 0.977 | 0.971 | 0.971 | 0.975 | 0.964 |
同环境复现 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
公开预训练权重的融合版本(不预处理) | 0.975 | 0.957 | 0.978 | 0.946 | 0.952 | 0.944 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
双库版本(processed)
双库版本(raw)
预训练uff日志
原参数第8个epoch的结果
原参数第3个epoch的结果
原参数第17个epoch的结果
原参数第200个epoch的结果
原参数第400个epoch的结果
ㅤ | ㅤ | Bagel | Cable_Gland | Carrot | Cookie | Dowel | Foam | Mean |
epochs=3 | I-AUROC | 0.782 | 0.653 | ㅤ | ㅤ | ㅤ | ㅤ | 0.7175 |
ㅤ | P-AUROC | 0.971 | 0.956 | ㅤ | ㅤ | ㅤ | ㅤ | 0.9635 |
ㅤ | AUPRO | 0.894 | 0.859 | ㅤ | ㅤ | ㅤ | ㅤ | 0.8765 |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
epochs=8 | I-AUROC | 0.808 | 0.540 | 0.775 | 0.601 | 0.713 | 0.739 | 0.696 |
ㅤ | P-AUROC | 0.975 | 0.960 | 0.989 | 0.926 | 0.912 | 0.978 | 0.957 |
ㅤ | AUPRO | 0.920 | 0.861 | 0.958 | 0.804 | 0.772 | 0.914 | 0.872 |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
epochs=17 | I-AUROC | 0.876 | 0.592 | 0.730 | 0.650 | 0.691 | ㅤ | 0.708 |
ㅤ | P-AUROC | 0.985 | 0.958 | 0.989 | 0.913 | 0.928 | ㅤ | 0.955 |
ㅤ | AUPRO | 0.946 | 0.862 | 0.959 | 0.778 | 0.801 | ㅤ | 0.869 |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
epochs=200 | I-AUROC | 0.811 | 0.823 | 0.853 | 0.717 | 0.678 | 0.629 | 0.752 |
ㅤ | P-AUROC | 0.990 | 0.986 | 0.989 | 0.960 | 0.974 | 0.908 | 0.968 |
ㅤ | AUPRO | 0.960 | 0.953 | 0.960 | 0.888 | 0.911 | 0.699 | 0.896 |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
epochs=400 | I-AUROC | 0.756 | 0.683 | 0.723 | 0.771 | 0.638 | ㅤ | 0.714 |
ㅤ | P-AUROC | 0.977 | 0.971 | 0.978 | 0.943 | 0.973 | ㅤ | ㅤ |
ㅤ | AUPRO | 0.921 | 0.926 | 0.918 | 0.854 | 0.910 | ㅤ | ㅤ |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
公开权重 | I-AUROC | 0.994 | 0.993 | 0.997 | 0.985 | 0.985 | 0.98 | ㅤ |
ㅤ | P-AUROC | 0.994 | 0.993 | 0.997 | 0.985 | 0.985 | 0.98 | ㅤ |
ㅤ | AUPRO | 0.969 | 0.97 | 0.978 | 0.948 | 0.94 | 0.918 | ㅤ |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | ㅤ | Bagel | Cable_Gland | Carrot | Cookie | Dowel | Foam | Mean |
epochs=199 | I-AUROC | 0.780 | 0.646 | 0.488 | 0.472 | ㅤ | ㅤ | ㅤ |
ㅤ | P-AUROC | 0.977 | 0.945 | 0.975 | 0.918 | ㅤ | ㅤ | ㅤ |
ㅤ | AUPRO | 0.911 | 0.832 | 0.914 | 0.755 | ㅤ | ㅤ | ㅤ |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
epochs=8 | I-AUROC | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | P-AUROC | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | AUPRO | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
epochs=17 | I-AUROC | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | P-AUROC | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | AUPRO | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
epochs=30 | I-AUROC | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | P-AUROC | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | AUPRO | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
BTF(3D-ADS)——RTX3090
日志(不预处理)
BTF I-AUROC
ㅤ | Bagel | Cable_Gland | Carrot | Cookie | Dowel | Foam | Peach | Potato | Rope | Tire | Mean |
论文报告 | 0.938 | 0.765 | 0.972 | 0.888 | 0.960 | 0.664 | 0.904 | 0.929 | 0.982 | 0.726 | 0.873 |
同环境复现 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
不预处理 | 0.951 | 0.811 | 0.975 | 0.888 | 0.961 | 0.587 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
BTF P-AUROC
ㅤ | Bagel | Cable_Gland | Carrot | Cookie | Dowel | Foam | Peach | Potato | Rope | Tire | Mean |
论文报告 | 0.996 | 0.991 | 0.997 | 0.995 | 0.995 | 0.972 | 0.996 | 0.998 | 0.995 | 0.994 | 0.993 |
同环境复现 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
不预处理 | 0.995 | 0.992 | 0.997 | 0.995 | 0.995 | 0.973 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
BTF AUPRO
ㅤ | Bagel | Cable_Gland | Carrot | Cookie | Dowel | Foam | Peach | Potato | Rope | Tire | Mean |
论文报告 | 0.976 | 0.967 | 0.979 | 0.974 | 0.971 | 0.884 | 0.976 | 0.981 | 0.959 | 0.971 | 0.964 |
同环境复现 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
不预处理 | 0.975 | 0.967 | 0.979 | 0.975 | 0.972 | 0.877 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
EasyNet
EasyNet
TaoTao9 • Updated Jun 20, 2024
DRAEM
VitjanZ • Updated Jul 17, 2024
回滚两次 恢复训练文件和readme
训练bagel日志
ㅤ | ㅤ | Bagel | Cable_Gland | Carrot | Cookie | Dowel | Foam | Mean |
论文数值 | I-AUROC | 0.991 | 0.998 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | P-AUROC | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | AUPRO | 0.839 | 0.864 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
复现数值(250epoch) | I-AUROC | 0.960 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | P-AUROC | 0.990 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | AUPRO | 0.908 | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ | ㅤ |
🤗 总结归纳
📎 参考文章
- 作者:ziuch
- 链接:https://ziuch.com/article/Point-cloud-preprocessing-paper-material
- 声明:本文采用 CC BY-NC-SA 4.0 许可协议,转载请注明出处。
相关文章