Avoid overflow in addition of bspline offset for integer-type converters #179
Conversation
to avoid undefined symbol errors in certain uses of the library
to allow access for pickling in pyapr
to avoid overflows when the input image has values near the largest representable number. New behavior is that no offset is added when either a) `lambda<=0` or b) the image covers the entire range of values.
required for `FetchContent_MakeAvailable` when building gtest
If not externally available, gtest is downlaoded and built using CMake's FetchContent
|
So the current failing test seem to be down to: In the sphere2D case; the local intensity scale is exactly 0. This results in nan's, it seem like all other systems this results in 0's for the the later level function. However, in macosx, it seems this now behaves differently. Tests passing can be recovered by adding 1 to the 0 entries in the LIS. I am reluctant to add another operation in main, to all paths to handle this case without thought. @joeljonsson what do you think? |
|
So it seemed the thesholds on this example were set to 0's. Which should never really be the case, but did not seem to matter up to this point. |
|
It seems that all test files are computed with As discussed before we should also add a warning, e.g. here, when both |
|
Added the warning and a test. To be honest; changing all the test files seems more dangerous, let along more work; then this current fix for now. |
| project(APR DESCRIPTION "Adaptive Particle Representation library") | ||
|
|
||
| message(STATUS "CMAKE VERSION ${CMAKE_VERSION}") | ||
|
|
||
| set(CMAKE_CXX_STANDARD 14) | ||
| set(CMAKE_CXX_STANDARD_REQUIRED ON) | ||
|
|
||
| if(POLICY CMP0135) |
There was a problem hiding this comment.
Why was this needed to be added?
There was a problem hiding this comment.
Honestly I'm not sure, but I think it has to do with the way gtest is now downloaded
Resolves #178
The bspline offset is now computed based on the range of values in the input image, and only if smoothing is enabled (
lambda>0).Small revision of examples:
Example_get_apr*now use converters of typefloat.-auto_parametersoption toExample_get_apr_by_blockExample_reconstruct_imageis now controlled via the option-noise <sigma>where the standard deviation can be set (default 0 = no additional noise)