New correspondence rejector (#940)

[andersgb1]

Implementation of the proposed feature here: http://dev.pointclouds.org/issues/940

[aichim]

Awesome, Anders! Will have a look soon and merge / comment.

Cheers,
Alex

[andersgb1]

Great, thanks Alex!

[aichim]

Hi Anders,

I went through your code and did some minor beautifications. The problem is that when I tried running the unit tests, I get this (on a 64-bit Mac OS):

/Users/alex/Work/pointclouds/pcl_github/pcl/test/registration/test_correspondence_rejectors.cpp:149: Failure
The difference between precision and 1.0 is 0.3666666666666667, which exceeds 0.35, where
precision evaluates to 0.6333333333333333,
1.0 evaluates to 1, and
0.35 evaluates to 0.34999999999999998.

Not sure how bad it is.

Thanks,
Alex

[andersgb1]

Ahh, I see. That's easily solved by changing line 149 in test_correspondence_rejectors from:

EXPECT_NEAR(precision, 1.0, 0.35);

to:

EXPECT_NEAR(precision, 1.0, 0.4);

The thresholds are purely constructed by the test case, so it's just a simple matter of varying machine precisions. Will you do it?

[aichim]

I can do it, sure, but isn't this going to be a problem for the functioning of the algorithm on different platforms?

[andersgb1]

I can (almost) guarantee you it will not :) The precision problem lies in the thresholds that I calculate inside the test code. The algorithm outputs the same absolute (integer) number of correspondences after execution and should definitely not be affected.

[aichim]

Done, thanks Anders!

[andersgb1]

Thank you, Alex!