Pass the arrays by reference rather than by value,
should give some percent of speedup.
Also don't pass the dimensions to the function but
get them from the images themselves.
Hopefully this will give some %% of tracker speedup.
This means if one makes change to either SConscript or CMakeLists.txt
there he MUST update bundle.sh.
Also made it so *.cc files from intern/libmv are matching which would
make it easier to add more .cc files there if needed.
And one more thing is removing 'simple_pipeline/distortion_models.cc'
which doesn't match any of the files.
This commit makes it so CameraIntrinsics is no longer hardcoded
to use the traditional polynomial radial distortion model. Currently
the distortion code has generic logic which is shared between
different distortion models, but had no other models until now.
This moves everything specific to the polynomial radial distortion
to a subclass PolynomialDistortionCameraIntrinsics(), and adds a
new division distortion model suitable for cameras such as the
GoPro which have much stronger distortion due to their fisheye lens.
This also cleans up the internal API of CameraIntrinsics to make
it easier to understand and reduces old C-style code.
New distortion model is available in the Lens panel of MCE.
- Polynomial is the old well-known model
- Division is the new one which s intended to deal better with huge
distortion.
Coefficients of this model works independent from each other
and for division model one probably want to have positive values
to have a barrel distortion.
This gives a huge speedup gain for cases when you've got
rather huge markers on a byte images.
Done by skipping IMB_float_from_rect()/IMB_rect_from_float()
for such cases. We can sample the buffers without color space
conversion.
This was a nasty bug which was caused by specific of how face-edge
attributes are stored in Carve.
Face pointer is used in the map key which works just fine in all
cases except for the cases when some face is getting freed after
it was stored in the map.
This might give real issues when new face is allocating because
it's possible new face would have the same address as the freed
one.
Such cases used to happen when union of separate manifolds is
needed for the operands AND jemalloc is enabled.
Solved by dropping attributes for the freed faces from the map.
Maybe not the fastest ever approach, but not sure how to make
it faster actually. Should work just fine. It only happens for
complex setups with intersecting manifolds in the operands.
-ffast-math is evil, not sure why it was enabled...
I seems to work better on OSX but it's still not a good idea.
The SConscript for bullet is a mess, I don't understand why
we use different flags for different platforms in the first place.
Seems to be a historical artifact but I don't know enough about scons
to try and clean it up.
Using unordered_map and unordered_set C++ container types currently
requires careful testing or usage of boost, due to the various confusing
C++ version differences in include paths and namespaces.
Libmv defines tests for these cases in cmake and scons, such that ceres
can use any available implementation, or fall back too std::map/std::set
if none can be found.
This patch generalizes this buildfile code by providing a Blender macro.
* cmake: defines both the variables used by libmv at them moment as well
as 2 variables UNORDERED_MAP_INCLUDE_PREFIX and UNORDERED_MAP_NAMESPACE,
which can later be used in other C++ parts for convenience.
* scons: adds a tool script returning the include prefix and namespace.
Libmv checks these to define the appropriate definitions for ceres.
Differential Revision: https://developer.blender.org/D425
This is mainly a maintaince commit which syncs changes
between Blender and Libmv upstream also bringing new
GLog version.
This GLog version is presumably have better support of
MinGW from "the box".
This commit is also aimed to make further 3d part libs
update easier.
This re-applied patch from 25cbd13 which was lost at some point
since missing patchset in series.
This revision is to be back-ported to the final release.
I overlooked in cmake my fix same time changed the optimization level to 0,
so not use optimization is the real fix ( postponed for after 2.70 ).
I appears we should investigate scons anyway: compileflags does not apply to c and c++ same time as expected.
This was caused by static variables used in plNearestPoints().
For now solved by making the solvers allocated in the stack,
seems no noticeable affect on the simulation speed so far.
There were loads of issues in the code still which are mow likely fixed:
- Hole resolver hook had memory leak -- it didn't free face with holes
when triangulating it.
- Original edge mapping didn't work correct. old code related on the fact
that loop order is not changing when constructing the MeshSet class, but
in fact it does change.
Currently used edge map for this because it was easiest way to do it now,
but after the release we're to change it. Main reason is that face mapping
is not correct as well (and it was never correct actually). So we'll need
to construct Mesh structures by our own to be sure we're using correct
original index mapping.
- Carve might produce faces with ears, which is forbidden in Blender.
it wasn't an issue in old integration because triangulation will remove
the ears. So for now simply added ears removing back as a hook.
But actual reason of the ears is to be investigated really.
This hook will only work for NGons, quads are assumed not be able to
have ears. So this additional hook shouldn't slow down things much.
- Carve's hole resolver produces duplicated faces in some cases. Still not
sure what is the reason of this. Code here is not so much straightforward,
this is to be investigated later.
For now solved the issue as own hole resolver which checks for duplicated
faces after the hole resolving.
The additional checks here will only run if the mesh actually have hole
and wouldn't introduce slowdown for faces which doesn't have holes.
- Made it so if edge user triangulation gets a split (for example, in cases
when this edge intersects with the second operand) it wouldn't be dissolved.
This prevents cases of crappy topology after dissolving in several cases.
- Edge dissolver didn't check for whether edge is a non-manifold. We couldn't
really dissolve open manifold edges.
The bad thing about this is that mesh triangulation might produce non-manifold
edges and they wouldn't be dissolved. Not worst case in the world, but would
be nice to have it solved somehow.
- Exporting mesh form Carve to Blender might have produced duplicated edges
in cases when several non-manifold faces shared the edge. This is also fixed
now.
- Mesh triangulation might have produced duplicated faces, which is really bad.
Fixed by keeping a track on which faces we've created and skipping adding new
triangle if we already have one.
This all might introduce some slowdown, but we're too close to the release now,
so would rather have it slower but robust. After the release we might look into
ways to speed things up.
This is needed to minimize their reprojection error over the footage.
Without this extra step positions of such tracks were calculated by
algebraic intersection code only, which doesn't give best precision.
Avoid zero-sized problem when doing euclidean intersection
Zero-sized problem might occur when intersecting track with
constant zero weight. For such tracks we'll just use result
of algebraic intersection.
TODO: We probably need to have a separate BA step to adjust
positions of tracks with constant zero weight.
It doesn't make sense to use zero-weighted tracks as a correspondences
in keyframe selection.
Such tracks are not guaranteed to be tracked accurately because their
purpose is to add reference points in 3D space without affecting the
solution.
This gives much worse results on mango footage (see 04_2e)
so disabling for now for until proper prediction model is landed.
The thing is, currently blender sends input coordinates as the guess to
region tracker and in case of fast motion such an early out ruins the track.
