Hello @arnaudbrejeon thanks for the bug report and fix!

@arnaudbrejeon Well done! Thanks for the reproducer!

Reference on original patch: #8789 (#7705).

If h is odd

Your guess seems correct.

Perhaps "odd" case is not expected at all by processing code:

• chunksSizeAndLabels_[r] = range.end;
• chunksSizeAndLabels_[startR + 1] = label - firstLabel;

and finally:

• const double nParallelStripes = std::max(1, std::min(h / 2, getNumThreads()*4));
• with L.rows / nThreads >= 2 limit:

    //Run parallel labeling only if the rows of the image are at least twice the number of available threads
    const bool is_parallel = currentParallelFramework != NULL && nThreads > 1 && L.rows / nThreads >= 2;

Last statements trying to guarantee that subtask size is not less than 2 rows (or may be even "even").
Anyway this approach is wrong because nStripes parameter of parallel_for() is just a hint and it is not accurate at all (like any result of floating-point calculations).

There is similar problem with 4 connectivity stripes: (r * imgLabels_.cols + 1) / 2 + 1

• r=3, cols=3 => 6
• r=4, cols=3 => 7
• r=6, cols=3 => 10
• 1 0 1 => 2 labels are required for 3 cols x 1 row
• 1 0 1 | 0 1 0 | 1 0 1 => up to 5 labels for 3 x 3 stripe (but 10-6=4)

Q: Why are we observing crash instead of incorrect result?

A: Uninitialized array with garbage after the first element:

            //Tree of labels
            LabelT *P = (LabelT *)cv::fastMalloc(Plength * sizeof(LabelT));
            //First label is for background
            P[0] = 0;

There is one more similar place in LabelingGranaParallel (8 connectivity):

LabelT label = LabelT((r + 1) / 2)  * LabelT((imgLabels_.cols + 1) / 2) + 1;

Additional commits will be pushed shortly.