doc/tutorials/features2d/akaze_matching/akaze_matching.markdown

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