Fix Akaze's rotational invarience. #64
Conversation
| @@ -318,10 +318,13 @@ fn do_subpixel_refinement( | |||
| if f32::abs(dst[0]) <= 1.0 && f32::abs(dst[1]) <= 1.0 { | |||
| let mut keypoint_clone = *keypoint; | |||
| keypoint_clone.point = ((x as f32) + dst[0], (y as f32) + dst[1]); | |||
| let power = f32::powf(2.0f32, evolutions[keypoint.class_id].octave as f32); | |||
There was a problem hiding this comment.
This change leads to points being out of bounds later, but it is what the C++ code does. I believe that either there is a matching bug somewhere else, or the C++ code might have an illegal memory access that doesn't get noticed due to the use of pointers.
There was a problem hiding this comment.
After putting an assertion in the C++ original code, I can confirm that it doesn't do out of bounds access where the Rust code does, so there is an additional bug somewhere.
|
That test error is just checking to ensure the behavior stays the same when changes are made. I can update the tests to match the new behavior. |
|
@stephanemagnenat I invited you as a member to Rust CV. Once you accept, you can modify this branch in the upstream repo. Since this PR didn't have the "allow maintainers to edit your branch" checked, it is easier for me to modify it here. I will also be adding changes to fix the other issues I identified in #63 which you can test against. Closing this for now, replaced with #65. |
Work in progress fixing the rotational invariance bug of Rust CV Akaze implementation.