Implements an automatic keyframe selection algorithm which uses
couple of approaches to find out best keyframes candidates:
- First, slightly modifier Pollefeys's criteria is used, which
limits correspondence ration from 80% to 100%. This allows to
reject keyframe candidate early without doing heavy math in
cases there're not much common features with first keyframe.
- Second step is based on Geometric Robust Information Criteria
(aka GRIC), which checks whether features motion between
candidate keyframes is better defined by homography or
fundamental matrices.
To be a good keyframe candidate, fundamental matrix need to
define motion better than homography (in this case F-GRIC will
be smaller than H-GRIC).
This two criteria are well described in this paper:
http://www.cs.ait.ac.th/~mdailey/papers/Tahir-KeyFrame.pdf
- Final step is based on estimating reconstruction error of
a full-scene solution using candidate keyframes. This part
is based on the following paper:
ftp://ftp.tnt.uni-hannover.de/pub/papers/2004/ECCV2004-TTHBAW.pdf
This step requires reconstruction using candidate keyframes
and obtaining covariance matrix of 3D points positions.
Reconstruction was done pretty much straightforward using
other simple pipeline routines, and for covariance estimation
pseudo-inverse of Hessian is used, which is in this case
(J^T * J)+, where + denotes pseudo-inverse.
Jacobian matrix is estimating using Ceres evaluate API.
This is also crucial to get rid of possible gauge ambiguity,
which is in our case made by zero-ing 7 (by gauge freedoms
number) eigen values in pseudo-inverse.
There're still room for improving and optimizing the code,
but we need some point to start with anyway :)
Thanks to Keir Mierle and Sameer Agarwal who assisted a lot
to make this feature working.
Will be reverted as soon as the x64 compiler is fixed.
For now it shouldn't have an impact on tracking performance. My test have shown no significant speed difference to official VC2008 build of 2.67.
- Added const modifiers where it makes sense and
helps keep code safe.
- Reshuffled argument to match <inputs>,<outputs>
convention on parameters.
- Pass values to ApplyRadialDistortionCameraIntrinsics
by a constant reference.
This will save lots of CPU ticks passing relatively
heavy jet objects to this function when running
bundle adjustment.
Makes code in tracking.cc much easier to understand and modify,
without worring to breck compulation with Libmv disabled.
It is still possible compilation will break due to libmv-capi
changes, but that's not happening so much often.
- Ensures fix for msvc2012 is applying correct.
- Some code cleanup to match libmv's code style.
- Do not include points which were intersect
behind the camera to a reconstruction.
- Includes changes needed for keyframe selection.
Additional changes:
- Cleaned up sources to reduce mess in some
big functions.
- Removed unused function from libmv c-api.
- Made functions naming more consistent.
- Use bool for internal stuff in tracking.c.
Shall be no functional changes :)
This check is actually redundant, because empty intrinsics
will have focal length of 1.0, which means original comment
about BundleIntrinsics was not truth.
It is possible that external user will send focal length of
zero to be refined, but blender prevents this from happening.
In cases keyframes are no so good, algebraic two frames construction
could produce result, for which more aggressive Ceres-based BA code
will fall to a solution for which points goes behind the camera,
which is not so nice.
Seems in newer Ceres returning false from cost functor wouldn't
abort solution, but will restrict solver from moving points behind
the camera.
Works fine in own tests, but requires more tests.
Made it so reconstructed scene always scaled in a way
that variance of camera centers is unity.
This solves "issues" when different keyframes will
give the same reprojection error but will give scenes
with different.scale, which could easily have been
considered as a bad keyframe combination.
This change is essential for automatic keyframe
selection algorithm to work reliable for user.
In some cases (was noticed on not good enough keyframe
pair) bundle adjuster could have moved bundles behind
the camera.
This could indeed lead to lower rewprojection error but
this is just pointless thing to do.
Now added check to residuals functor which will return
false to Ceres in cases point moved behind the camera
to prevent such issues.
This brings a fixes for threading issue in BLAS
making BA step more robust (there were some in-detemrinacy
caused by this threading issue).
Also brings some optimizations, which does not directly
affect on blender.
This commit bundles new libmv version from own branch
which brings fix for wrong parameter block used for
modal solver parameterization.
Fixes#34985: Crash with Motion tracker (Tripod Motion)
- Get rid of rotation matrix parameterization,
use angle-axis instead.
Also Joined rotation and translation into a
single parameter block.
This made minimization go significantly faster,
like 1.3x times in average.
- Fix first camera when bundling. This is to
address orientation ambiguity.
Reconstruction result could still vary in
size, but that's another issue to be addressed
later.
Additional change:
Split EuclideanBundleCommonIntrinsics into
smaller functions, so it's now a bit easier
to follow.
This made preview working but that broke internals
of tracking.
Namely, BlurredImageAndDerivativesChannels is giving
much more blurred image because it was assuming pixel
center is an integer position.
Guess other parts of libmv used to suffer because of
this issue.
Now pixel centering happens in blender side, and
libmv assumes integer position is a pixel center.
Thins brings up some speed improvements:
SPARSE_SCHUR is approx 1.3-1.5x times faster
ITERATIVE_SCHUR is approx 1.2x times faster
For blender this means camera solution go a bit
faster now. Would not have affect on tracking
speed.
This commit implements multi-threaded calculation of frames
when building proxies. Both scaling and undistortion steps
are now threaded.
Frames and proxy resolution are still handled one-by-one,
saving files after every single step. So if HDD is not so
fast, this commit could have not so much benefit.
Internal changes:
- Added IMB_scaleImBuf_threaded which scales given image
buffer in multiple threads and uses bilinear filtering.
- libmv's camera intrinsics now have SetThreads() method
which is used to specify how many OpenMP threads to use
for buffer distortion/undistortion.
And yeah, this code is using OpenMP for threading.
- Reshuffled a bit libmv-capi calls and added function
BKE_tracking_distortion_set_threads to specify number
of threads used by intrinscis.
There're some features planned which would
require rigid registration, but this code
would need to be re-done anyway to use new
minimizer and solving some issues with ICP
algorithm there.
Several major things are done in this commit:
- First of all, logic of modal solver was changed.
We do not rely on only minimizer to take care of
guessing rotation for frame, but we're using
analytical rotation computation for point clouds
to obtain initial rotation.
Then this rotation is being refined using Ceres
minimizer and now instead of minimizing average
distance between points of point of two clouds,
minimization of reprojection error of point
cloud onto frame happens.
This gives quite a bit of precision improvement.
- Second bigger improvement here is using bundle
adjustment for a result of first step when we're
only estimating rotation between neighbor images
and reprojecting markers.
This averages error across the image sequence
avoiding error accumulation. Also, this will
tweak bundles themselves a bit for better match.
- And last bigger improvement here is support of
camera intrinsics refirenment.
This allowed to significantly improve solution
for real-life footage and results after such
refining are much more usable than it were before.
Thanks to Keir for the help and code review.
This information is useful, but in cases when you, say,
working on a bundler it's annoying to scroll all the
information up.
Now behavior would be:
- running `./blender --debug-libmv` will print all the
debug messages
- running `./blender --debug-libmv --verbose 0` will
print only debug messages from solvers, recosntruction
and so, but will bypass final reprojection bunch of
messages
- running `./blender --debug-lib,v --verbose 1` will
include final reprojection messages.
This shall not lead to any functional changes, just
avoids radial distortion code duplicated in camera
intrinsics and bundling code.
For fancier bundle adjustment supprting different
distortion models this is not actually enough and
would need to make some bigger changes, but this
changes makes code a bit easier to maintain already.
Also made it theraded linear solver, seems it makes
sense for iterative schur with inner iterations
enabled.
Use OpenMO's max therads called from bundler code
to detect how many threads to use. Could be changed
in a way that number of threads is passing in options
from blender side in the future.
Also removed redundant V3D definition from compiler's
flags.
This commits adds extra refirenment entry in the menu which is
"K1, K2" and which will apparently refine only this distortion
coefficients.
This would be useful in cases when you know for sure focal length
(which could be obtained from lens, EXIF and so) but not sure
about how good you manual calibration is.
Be careful tho, there're no internal constraints on this
coefficients so distortion model could just screw up into insane
values.
Would not expect any significant changes in solver behavior, but
it could be more accurate in some cases.
Switching projective intersection to ceres is marked as a TODO
for now.
With new bundle adjustment based on Ceres we don't need
SSBA library anymore. This also means we don't need ldl
library and libmv is no longer depends on colamd as well.
Patch originally written by me, then finished by Sergey. Big
thanks to Sergey for troopering through and fixing the many issues
with my original (not compilable) patch.
The Ceres implementation uses 2 parameter blocks for each camera
(1 for rotation and 1 for translation), 1 parameter block for
common intrinsics (focal length etc) and 1 parameter block for
each track (e.g. bundle or 3D point).
We turn on some fancy optimizer options to get better performance,
in particular:
options.preconditioner_type = ceres::SCHUR_JACOBI;
options.linear_solver_type = ceres::ITERATIVE_SCHUR;
options.use_inner_iterations = true;
options.use_nonmonotonic_steps = true;
options.max_num_iterations = 100;
Special thanks to Sameer Agarwal of Ceres fame for splitting out
the SCHUR_JACOBI preconditioner so that it didn't depend on
CHOLMOD. Previously we could not use that preconditioner in
Blender because CHOLMOD is too large of a dependency for Blender.
BundleIntrinsicsLogMessage:
- Moved bunch of if(foo) LG << "bar" into this function, to make
EuclideanBundleCommonIntrinsics a little bit easier to follow.
EuclideanBundle:
- Fix RMSE logging.
structure instead of passing all the parameters to every function.
Makes it much easier to tweak distortion model.
---
svn merge -r52854:52855 ^/branches/soc-2011-tomato
