Skip to content

linyicheng1/keypoint_bench

BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

43 Commits
config
config
 
 
datasets
datasets
 
 
models
models
 
 
script
script
 
 
tasks
tasks
 
 
utils
utils
 
 
weights
weights
 
 
Dockerfile
Dockerfile
 
 
README.md
README.md
 
 
main.py
main.py
 
 
requirements.txt
requirements.txt
 
 
run.sh
run.sh
 
 

Repository files navigation

图像点特征性能指标对比

支持的点特征

  • ORB
  • SIFT
  • Harris
  • SuperPoint
  • ALIKE
  • XFeat
  • D2-Net
  • DISK
  • R2D2
  • SFD2

支持的匹配方法

  • 暴力匹配
  • lightglue
  • 光流

评价指标

  • AUC 基于 MegaDepth 数据集
  • MHA 基于 Hpatch 数据集
  • 重复率 基于 Hpatch 数据集
  • 基础矩阵 基于TartanAir数据集
  • VO里程计 基于KITTI数据集

RUN IN Docker

  1. 构造docker镜像
mkdir keypoint_bench && cd keypoint_bench
wget https://github.com/linyicheng1/keypoint_bench/blob/main/Dockerfile
docker build -t keypoint_bench:v1 .

需要等待10-20分钟，下载依赖并且构建镜像

  1. 运行docker容器
docker run -it --gpus all -u root -p 2223:22 -v /数据集下载位置/:/home/data/ keypoint_bench:v1
  1. 运行python程序

在运行程序前需要修改配置文件中包含的数据集路径，按自己的实际情况修改/home/code/keypoint_bench/config/config.yaml文件

运行程序

cd /home/code/keypoint_bench
python3 main.py --config_file config/config.yaml test

默认得到ALIKE的重复率计算结果

repeatability 0.3157695 rep_mean_err 1.2313193

一些结果

参数

值得注意的是，许多论文为了指标更高，往往提取更多的特征，并且非极大值抑制范围很小，这样在实际使用的时候会造成特征扎堆的现象。因此，我们使用的参数都是特征点数量较少，并且非极大值抑制的窗口比较大，这更符合实际情况。尽管这样往往会导致指标较低。

特征提取参数
  • nms_dist: 6
  • min_score: 0.0
  • top_k: 1000
  • threshold: 0
  • border_dist: 8
暴力匹配参数
  • metric: euclidean
  • max_distance: 5
  • cross_check: True
光流匹配参数
  • distance: 10
  • win_size: 21
  • levels: 3
  • interation: 40

结果

重复率
特征点 Harris ALIKE SuperPoint XFeat D2-Net DISK R2D2 SFD2 LGood (Ours) EdgePoint (Ours)
重复率 0.204 0.319 0.359 0.177 0.179 0.290 0.331 0.381 0.402 0.349
平均误差 1.861 1.317 1.466 1.508 2.022 1.473 1.483 1.507 1.447 1.390
MHA (暴力匹配)
特征点 ALIKE SuperPoint XFeat D2-Net DISK R2D2 SFD2 EdgePoint (Ours)
MHA@3 0.491 0.437 0.372 0.016 0.254 0.335 0.467 0.461
MHA@5 0.583 0.569 0.511 0.105 0.410 0.488 0.572 0.572
MHA@7 0.657 0.613 0.589 0.220 0.525 0.555 0.620 0.640
AUC (暴力匹配) MegaDepth
特征点 ALIKE SuperPoint XFeat D2-Net DISK R2D2 SFD2 EdgePoint (Ours)
AUC@5 0.375 0.192 0.200 0.126 0.249 0.247 0.295 0.304
AUC@10 0.527 0.323 0.351 0.240 0.403 0.360 0.442 0.448
AUC@20 0.661 0.461 0.516 0.386 0.562 0.464 0.578 0.581
基础矩阵 (光流跟踪) TartanAir

误差

特征点 ALIKE SuperPoint XFeat D2-Net DISK R2D2 SFD2 EdgePoint (Ours) LGood (Ours) LET-NET (Ours) Harris
MH000 9.559 9.441 9.609 9.868 9.466 9.888 9.355 9.712 7.374 9.015 9.585
MH001 11.460 11.719 12.201 11.004 9.470 11.784 11.738 11.260 7.872 9.774 11.376
MH002 9.459 9.480 9.947 9.789 9.069 9.301 9.534 9.372 7.357 7.855 9.796
MH003 14.128 13.837 14.505 12.925 12.871 13.966 13.831 13.573 8.649 10.065 13.607
MH004 21.936 22.745 22.466 19.999 20.051 21.898 22.115 20.900 10.554 13.167 21.654
MH005 7.856 7.895 7.879 7.855 7.093 7.846 7.870 7.996 7.213 7.158 7.778
MH006 13.233 12.846 13.555 12.308 12.698 13.111 13.145 12.763 8.304 10.052 12.838
MH007 10.612 10.599 10.861 10.438 9.789 10.659 10.644 10.472 7.919 9.182 10.618
ME000 9.428 9.663 9.524 10.130 8.842 9.795 9.380 9.319 6.540 7.896 9.301
ME001 6.165 6.602 6.428 6.705 5.285 6.381 6.521 6.206 6.166 5.980 6.468
ME002 7.383 6.583 6.934 6.423 5.832 6.563 6.487 6.535 6.055 6.154 6.514
ME003 8.887 8.534 8.658 8.283 7.979 8.819 8.588 8.444 7.293 7.829 8.611
ME004 8.534 8.331 8.551 8.911 7.346 8.601 8.386 8.717 7.213 7.762 8.241
ME005 12.486 14.145 13.886 9.450 10.215 12.963 11.536 12.775 6.600 9.984 11.732
ME006 6.609 6.785 6.684 7.007 6.223 6.675 6.779 6.757 6.387 6.331 6.478
ME007 9.732 9.643 9.805 9.372 9.022 9.975 9.705 9.629 7.436 8.74 9.878
基础矩阵 (描述子匹配) TartanAir
特征点 ALIKE SuperPoint XFeat D2-Net DISK R2D2 SFD2
MH000 104.336 31.485 104.114 0.395 230.685 202.505 6.030
MH001 171.056 49.141 121.606 0.392 236.237 262.874 12.700
MH002 178.670 18.793 118.860 0.425 233.026 256.604 5.109
MH003 101.301 33.025 86.855 0.337 186.509 178.325 4.344
MH004 98.713 7.307 98.488 0.376 243.231 178.595 4.478
MH005 246.712 50.350 144.445 0.390 226.917 270.599 9.568
MH006 163.939 42.238 120.138 0.373 210.018 233.142 8.151
MH007 171.425 43.656 127.893 0.393 245.529 248.681 8.436
ME000 121.683 36.422 114.715 0.501 331.300 256.842 9.606
ME001 243.317 38.217 139.288 0.393 217.231 254.585 10.482
ME002 214.653 66.151 137.479 0.449 253.256 283.337 13.906
ME003 122.109 24.904 85.554 0.340 196.310 189.215 7.827
ME004 86.689 38.143 131.865 0.478 265.158 221.082 3.672
ME005 107.440 14.886 86.840 0.384 239.132 190.755 9.145
ME006 231.071 44.124 133.078 0.377 201.123 263.733 9.126
ME007 202.551 28.271 140.878 0.436 274.537 283.592 8.675

About

No description, website, or topics provided.

Resources

Readme
Activity

Stars

7 stars

Watchers

1 watching

Forks

1 fork
Report repository

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •  

Languages