doc/tutorials/features2d/akaze_matching/akaze_matching.markdown

   1 AKAZE local features matching {#tutorial_akaze_matching}
   2 =============================
   3
   4 Introduction
   5 ------------
   6
   7 In this tutorial we will learn how to use AKAZE @cite ANB13 local features to detect and match keypoints on
   8 two images.
   9 We will find keypoints on a pair of images with given homography matrix, match them and count the
  10 number of inliers (i.e. matches that fit in the given homography).
  11
  12 You can find expanded version of this example here:
  13 <https://github.com/pablofdezalc/test_kaze_akaze_opencv>
  14
  15 Data
  16 ----
  17
  18 We are going to use images 1 and 3 from *Graffiti* sequence of [Oxford dataset](http://www.robots.ox.ac.uk/~vgg/data/data-aff.html).
  19
  20 ![](images/graf.png)
  21
  22 Homography is given by a 3 by 3 matrix:
  23 @code{.none}
  24 7.6285898e-01  -2.9922929e-01   2.2567123e+02
  25 3.3443473e-01   1.0143901e+00  -7.6999973e+01
  26 3.4663091e-04  -1.4364524e-05   1.0000000e+00
  27 @endcode
  28 You can find the images (*graf1.png*, *graf3.png*) and homography (*H1to3p.xml*) in
  29 *opencv/samples/data/*.
  30
  31 ### Source Code
  32
  33 @add_toggle_cpp
  34 -   **Downloadable code**: Click
  35     [here](https://raw.githubusercontent.com/opencv/opencv/3.4/samples/cpp/tutorial_code/features2D/AKAZE_match.cpp)
  36
  37 -   **Code at glance:**
  38     @include samples/cpp/tutorial_code/features2D/AKAZE_match.cpp
  39 @end_toggle
  40
  41 @add_toggle_java
  42 -   **Downloadable code**: Click
  43     [here](https://raw.githubusercontent.com/opencv/opencv/3.4/samples/java/tutorial_code/features2D/akaze_matching/AKAZEMatchDemo.java)
  44
  45 -   **Code at glance:**
  46     @include samples/java/tutorial_code/features2D/akaze_matching/AKAZEMatchDemo.java
  47 @end_toggle
  48
  49 @add_toggle_python
  50 -   **Downloadable code**: Click
  51     [here](https://raw.githubusercontent.com/opencv/opencv/3.4/samples/python/tutorial_code/features2D/akaze_matching/AKAZE_match.py)
  52
  53 -   **Code at glance:**
  54     @include samples/python/tutorial_code/features2D/akaze_matching/AKAZE_match.py
  55 @end_toggle
  56
  57 ### Explanation
  58
  59 -   **Load images and homography**
  60
  61 @add_toggle_cpp
  62 @snippet samples/cpp/tutorial_code/features2D/AKAZE_match.cpp load
  63 @end_toggle
  64
  65 @add_toggle_java
  66 @snippet samples/java/tutorial_code/features2D/akaze_matching/AKAZEMatchDemo.java load
  67 @end_toggle
  68
  69 @add_toggle_python
  70 @snippet samples/python/tutorial_code/features2D/akaze_matching/AKAZE_match.py load
  71 @end_toggle
  72
  73 We are loading grayscale images here. Homography is stored in the xml created with FileStorage.
  74
  75 -   **Detect keypoints and compute descriptors using AKAZE**
  76
  77 @add_toggle_cpp
  78 @snippet samples/cpp/tutorial_code/features2D/AKAZE_match.cpp AKAZE
  79 @end_toggle
  80
  81 @add_toggle_java
  82 @snippet samples/java/tutorial_code/features2D/akaze_matching/AKAZEMatchDemo.java AKAZE
  83 @end_toggle
  84
  85 @add_toggle_python
  86 @snippet samples/python/tutorial_code/features2D/akaze_matching/AKAZE_match.py AKAZE
  87 @end_toggle
  88
  89 We create AKAZE and detect and compute AKAZE keypoints and descriptors. Since we don't need the *mask*
  90 parameter, *noArray()* is used.
  91
  92 -   **Use brute-force matcher to find 2-nn matches**
  93
  94 @add_toggle_cpp
  95 @snippet samples/cpp/tutorial_code/features2D/AKAZE_match.cpp 2-nn matching
  96 @end_toggle
  97
  98 @add_toggle_java
  99 @snippet samples/java/tutorial_code/features2D/akaze_matching/AKAZEMatchDemo.java 2-nn matching
 100 @end_toggle
 101
 102 @add_toggle_python
 103 @snippet samples/python/tutorial_code/features2D/akaze_matching/AKAZE_match.py 2-nn matching
 104 @end_toggle
 105
 106 We use Hamming distance, because AKAZE uses binary descriptor by default.
 107
 108 -   **Use 2-nn matches and ratio criterion to find correct keypoint matches**
 109 @add_toggle_cpp
 110 @snippet samples/cpp/tutorial_code/features2D/AKAZE_match.cpp ratio test filtering
 111 @end_toggle
 112
 113 @add_toggle_java
 114 @snippet samples/java/tutorial_code/features2D/akaze_matching/AKAZEMatchDemo.java ratio test filtering
 115 @end_toggle
 116
 117 @add_toggle_python
 118 @snippet samples/python/tutorial_code/features2D/akaze_matching/AKAZE_match.py ratio test filtering
 119 @end_toggle
 120
 121 If the closest match distance is significantly lower than the second closest one, then the match is correct (match is not ambiguous).
 122
 123 -   **Check if our matches fit in the homography model**
 124
 125 @add_toggle_cpp
 126 @snippet samples/cpp/tutorial_code/features2D/AKAZE_match.cpp homography check
 127 @end_toggle
 128
 129 @add_toggle_java
 130 @snippet samples/java/tutorial_code/features2D/akaze_matching/AKAZEMatchDemo.java homography check
 131 @end_toggle
 132
 133 @add_toggle_python
 134 @snippet samples/python/tutorial_code/features2D/akaze_matching/AKAZE_match.py homography check
 135 @end_toggle
 136
 137 If the distance from first keypoint's projection to the second keypoint is less than threshold,
 138 then it fits the homography model.
 139
 140 We create a new set of matches for the inliers, because it is required by the drawing function.
 141
 142 -   **Output results**
 143
 144 @add_toggle_cpp
 145 @snippet samples/cpp/tutorial_code/features2D/AKAZE_match.cpp draw final matches
 146 @end_toggle
 147
 148 @add_toggle_java
 149 @snippet samples/java/tutorial_code/features2D/akaze_matching/AKAZEMatchDemo.java draw final matches
 150 @end_toggle
 151
 152 @add_toggle_python
 153 @snippet samples/python/tutorial_code/features2D/akaze_matching/AKAZE_match.py draw final matches
 154 @end_toggle
 155
 156 Here we save the resulting image and print some statistics.
 157
 158 Results
 159 -------
 160
 161 ### Found matches
 162
 163 ![](images/res.png)
 164
 165 Depending on your OpenCV version, you should get results coherent with:
 166
 167 @code{.none}
 168  Keypoints 1:   2943
 169  Keypoints 2:   3511
 170  Matches:       447
 171  Inliers:       308
 172  Inlier Ratio: 0.689038
 173 @endcode