martinRosevear's profile - activity

Dear All

I have found very rare problems with findContours() in x64. It causea an 'assert' failure when leaving the calling routine for some (rare) images.

The problem has been fixed (so-far) by re-coding the function into the 'C' version (cvFindContours()), and analyzing 50,000+ discrete images.

The cause is not immediatly obvious from the images which cause the failure, nor from the calling context.

Although the problem is currently solved, I do not like the thought of any lurking issues.

Any thoughts?

Martin

This should cover all the source files referenced by OpenCV routines. Currently many files referenced need to be supplied externally, covering extended instruction sets, Floating Point, parallel processing, TIFF file handling, among others, that are referenced when targeting the x86 and using the latest version of Visual Studio.

I have recently developed an OpenCV app on the PC (using the library) that worked well. I then tried porting that to the IoS and Android environment using the Marmalade SDK. This required compiling OpenCV source files using the GCC compiler for ARM/x86. This has proven difficult due to the large number of header files referenced by OpenCV code that need to be sourced externally when trying to debug in the x86 environment.

Hence one is forced to abandon the effort or write functions from scratch which largely defeats the purpose of using OpenCV.

In my understanding MSER.cpp detects both 'black on white' and 'white on black' MSERs. See routine extractMSER_8UC1() which calls both versions and can be modified with a bespoke compilation to detect only one type of MSER.

In my opinion, the main calling MSER procedure should be controlled by a flag to detect 'black on white', 'white on black' OR both types of MSERs.

Either way the documentation should be updated to make clear the performance and how it can be altered if a user requires it.