h2suzuki / fast_akaze Goto Github PK

Can you please make this run on visual studio 2017 and opencv 4.0? Or help me do it?

Reported by Pablo, the original author of A-Kaze.
https://gist.github.com/h2suzuki/ddc40e5913a1674ce6d9

There seems to be a standard deviation about 5% for the number of detected keypoints using the same image and different executions of your code.

AKAZE2 sporadically loses some descriptor matches.

$ ./akaze_match.exe img1.ppm img4.ppm H1to4p --descriptor 5
AKAZE_Impl2 constructor called
AKAZE_Impl2 constructor called
AKAZEFeaturesV2 constructor called
hardware_concurrency: 8
AKAZEFeaturesV2 constructor called
hardware_concurrency: 8
Number of Keypoints Image 1: 1822
Number of Keypoints Image 2: 1091
A-KAZE Features Extraction Time (ms): 120.327
Matching Descriptors Time (ms): 16.2663
Number of Matches: 959
Number of Inliers: 889
Number of Outliers: 70
Inliers Ratio: 92.7007

$ ./akaze_match.exe img1.ppm img4.ppm H1to4p --descriptor 5
AKAZE_Impl2 constructor called
AKAZE_Impl2 constructor called
AKAZEFeaturesV2 constructor called
hardware_concurrency: 8
AKAZEFeaturesV2 constructor called
hardware_concurrency: 8
Number of Keypoints Image 1: 1822
Number of Keypoints Image 2: 1091
A-KAZE Features Extraction Time (ms): 122.356
Matching Descriptors Time (ms): 16.2488
Number of Matches: 957  <--- OOPS TWO LESS!!
Number of Inliers: 889
Number of Outliers: 68  <--- OOPS TWO LESS!!
Inliers Ratio: 92.8945

Reported by Pablo, the original author of A-Kaze.

I have been working with the descriptor DESCRIPTOR_MLDB_UPRIGHT.
Since the the descriptor uses 486 bits, 61 bytes are needed and there are two bits that are not used.
In my opinion these two bits should be initialized to a fixed known value, for instance zero.
In fact when using descriptor DESCRIPTOR_MLDB, these initialization is done, but when using the upright version I have not seen where it is done in the code and have seen that these two extra bytes are random (what causes problems for matching)

I think that an easy solution would be to set line 1074 of AKAZEFeatures.cpp to
desc = 0;

Reported by Pablo, the original author of A-Kaze.
VLFeat Benchmarks shows that the original A-KAZE seems to perform a bit better.

It might be related to the change in R8.5: refactoring Compute_Main_Orientation().

Hi
I have just compiled the fast_akaze with OpenCV 3.4.5. Two things are strange:

1. fast_akaze (#define USE_AKAZE2 1) runs at 5-6 fps while OpenCV's akaze (#define USE_AKAZE2 0) runs at 11-12 fps!

2. Both versions are unable to do matching. I can see the keypoints with their size and orientation but no match is shown. Number of matches is always 0.