[Work Log] WACV reconstruction (revisited)
Addressing issues that came up last time we ran WACV.
Currently working on dataset #8, which has some real problems.
Spent some time tweaking the ground truth, adding missing curves, and correcting one or two obvious mistakes.
Still getting terrible results; one problem might be the "rough" reconstruction, which doesn't take into account anisotropic data uncertainty. Fixed.
False-positive reversal has been corrected, but still getting terrible results; it seems like the camera calibration is waaaay off.
I dont have the calibration data for this dataset available, locally re-syncing full dataset.
Are these calibrated cameras ordered in counter-clockwise direction? No. but the calibration_data.mat file in for this dataset is.
Inspected correspondence diagrams. Possibly some of the ground truth curves are reversed? Dataset #8, curve #7, view #9, only corresponds to very end of other points:
Tweaking ground truth tracing program to show start-point of curves.
Ground truth program doesn't compile. Clang++ for mountain lion is complaining.
Apparently clang no longer uses gnu stdc++ library, and some library components use proprietary gnu symbols (I admit, it was my code). Fixed.
Getting some new errors regarding some forward declarations of STL pair. Was hard to debug, because errors weren't local to the problematic code, and errors were cryptic. Removed forward declarations; fixed.
Found a template function I hand't ported from my experimental branch of KJB.
Linker errors -- libXmu and libXi not found. Tweaked init_compile's logic for OpenGL on Macs. Hopefully I didn't break anyone else's builds...
More linker errors. Apparently i need to recompile everything to use clang's c++ libraries?? rebuilt boost, next is casadi, but I can't find the source on my machine.
TODO (new)
- recalibrate cameras