Fix issues in Convex Hull algorithm #27020
Conversation
|
I created a simpler patch (than for #26975) that is considerably slower but seems to be correct. |
|
To fix this issue entirely i would propose something like the following: Preferred signature: easy to be implemented fast and correct. allows for input validation.template<typename TpointContainer, typename Thull>
void convexHull(TpointContainer& points, std::vector<Thull>& hull, bool clockwise>Deprecated signature 1: does the same as the preferred signature, only deprecated because the last parameter should be removed.template<typename TpointContainer, typename Tpoints, typename Thull>
void convexHull(TpointContainer& points, std::vector<Thull>& hull, bool clockwise, bool ignored>Deprecated signature 2: the original function fixed to the point where it retains behavior as closely as possible while preventing memory errors.
|
a408412 to
6ae1ea1
Compare
|
Afaics tests fail now when they are using numerical arrays as input. I wouldn't know how to fix that except by explicitly converting (i use _InputArray::copyTo now)... but actually arrays are not documented as possible input for convexHull. How should i proceed? |
|
There is no array kind. I am puzzled by how to make this work with all the variants. In short: how can i convert numerical arrays, std::vector<Point_> and cv::Matx_<Point_> via InputArray to a continuous memory region of points? how to distinguish them and perform input validation? int arr[2] = {2,3};
Point* pt = (Point*) arr;Just going through getMat() doesn't work for numerical arrays. and: how should i handle the discrepancies with the documentation? |
|
@mshabunin could you take a look? |
|
@mshabunin Could you take a look on it? |
|
I couldn't reproduce original issue, waiting for more details (see #26952). |
|
Sorry for the delay. I'll try to continue now but I don't know if i have enough time to finish. |
|
i am just asking because i wonder... you wouldn't ask for a reproducer if it was something like std::vector({0,1,2})[99] (and i think it is even worse). anyway i am on it. |
|
Friendly reminder. |
| for( i = 0; i < nout; i++ ) | ||
| *(Point*)(hull.ptr() + i*step) = data0[hullbuf[i]]; | ||
| } else { | ||
| _hull.create(nout, 1, _hull.type()); |
There was a problem hiding this comment.
_hull.type() may be undefined, e.g. in Python case of it output is just cv::Mat. The original approach works better here.
| for(int i = 0; i < bl_count-1; i++ ) | ||
| hullbuf[nout++] = int(&pointer[bl_stack[i]] - pointer); | ||
| for(int i = br_count-1; i > 0; i-- ) | ||
| hullbuf[nout++] = int(&pointer[br_stack[i]] - pointer); |
There was a problem hiding this comment.
&pointer[bl_stack[i]] - pointer is just bl_stack[i], right?
| for(int i = 0; i < tl_count-1; i++ ) | ||
| hullbuf[nout++] = int(&pointer[tl_stack[i]] - pointer); | ||
| for(int i = tr_count - 1; i > 0; i-- ) | ||
| hullbuf[nout++] = int(&pointer[tr_stack[i]] - pointer); |
There was a problem hiding this comment.
pointer[tl_stack[i]] - pointer is just tl_stack[i], right?
Pull Request Readiness Checklist
See: #26952 and #26975
Patch to opencv_extra has the same branch name.