System information (version)

• OpenCV => 3.1.0-7
• Operating System / Platform => 4.9.2-1-MANJARO x86-64
• Compiler => 6.3.1 20170109 (GCC)
• No OpenCl, Tegra or IPP

Detailed description

We wanted our opencv based source code to be compatible with both opencv 2.4 and 3.x. However, our program yielded unexpected results with 3.x. After some rigorous debugging we discovered the bug in Canny(). While the documentation states that the output matrix "can be the same as the input" this causes the problem. If we use a dedicated output matrix the results are as expected.

We assume that parallelization optimizations in 3.x causes this problem. Multiple threads operate read and write on the same matrix causing race conditions.

Steps to reproduce

```.cpp
Canny( detected_edges, detected_edges, lowThreshold, lowThreshold*ratio, kernel_size );
```

On a machine with multiple cpu cores

Workaround

```.cpp
Canny( detected_edges, tmp, lowThreshold, lowThreshold*ratio, kernel_size );
```