blender/extern/libmv/ChangeLog
Sergey Sharybin 8cca2fec82 Fix for bundle adjusting with motion restricted
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)
2013-04-14 23:43:00 +00:00

694 lines
24 KiB
Plaintext

commit 03cbc88ce7f51aa26ba503acea2e984bcb78873c
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Mon Apr 15 05:35:33 2013 +0600
Fix for bundle adjusting with motion restricted
Was a bug introduced in previous commit, which
was trying to set parameterization for non-existing
camera->t parameter block.
Replaced with subset parameterization.
Also added basic synthetic unit test for modal solver.
commit c78a68f980e778d40ce836fa1d7471cb7264d4a0
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Mon Apr 8 23:33:20 2013 +0600
Bundle adjustment improvements
- 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.
commit 28c7566629c2cf5b03a787c9509856e87472eb2f
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Mon Apr 8 23:31:57 2013 +0600
Update Ceres to current HEAD
Brings up some noticeable speed improvements. In particular
the automatic differentiation and bundle adjustment solvers.
commit efde9faa21e70b031d3cbcb2dcdcd38e597bf56e
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Mon Apr 8 02:21:26 2013 +0600
Corrected path to gflags
Currently tools/track.cc is not used, but let's
keep things a bit more up-to-date :)
commit f8b5ea196fb00ab07d577e9738a60cdd1de16509
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Mon Apr 8 02:17:16 2013 +0600
Re-enable tests for multiview and image
For as long code is in repo and used by some tools
better to have it covered by tests.
Some of them are failing tho, but that's completely
different story to be addressed later.
commit d2a7ee60a5845738f76b88bfc373eefc2cc8501a
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Mon Apr 8 02:10:07 2013 +0600
Do not modify cache's CMAKE_CXX_FLAGS_RELEASE when configuring Ceres
Otherwise you'll have infinite appending of Ceres-specific flags
on every saving of any CmakeLists.txt.
commit 45edb507bf46194dd55b7fc46a7d90ee3853834d
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Sun Apr 7 21:53:23 2013 +0600
Fixed compilation with BUILD_TOOLS enabled
This commit mainly reverts parts of following commits:
0eebc21db831211738acc938566bbc29d68d45db
d8109b7a4fede1660e0dbd73735f1a9e3fd79eec
e59595806c045916ab4ef15ef7047c1a728b2da9
2d6cd58ee1cd7c5073980f358c71b2898ad16b4c
They declared lots of stuff deprecated, but in
fact it's not deprecated just a bit different
usage pipeline. Anyway, deprecation shall not
happen spontaneously as a part of other changes.
And for sure shall not break anything.
commit 7a9c83d3ccaa2f0015f88b9156d7662c46244b4a
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Sat Apr 6 20:49:05 2013 +0600
Revert "Change libmv's bilinear sampling to assume the same"
Revert changes to bilinear sampler which were originally
aimed to match blender's pixel center (where integer coord
is a left-bottom corner, x.5 coords are centers.
The reason of revert is changing this assumption in only
sampler didn't work well and lead to wrong results of
BlurredImageAndDerivativesChannels for example.
Discovered when was doing unit-tests for brute region tracker.
This reverts commit daa354c0735b954b0cd7725626e9a3d67416d46b.
commit 15f3bb00340933ce753a1a55e9cde9383352e259
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Sat Apr 6 18:37:37 2013 +0600
Added basic test for brute region tracker
It is failing at this moment and this is caused because
of how SampleLinear works - seems it's assumption about
pixel center is not correct for internal sampling.
commit 1146602972c07e99a9e4ab37d35ac83aec490e60
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Sat Apr 6 16:54:08 2013 +0600
Tweak to KLT region tracker test
KLT is usually used to track relatively small
motions, and in this case motion almost equals
to half window size. This confuses math and
leads to not so much expected result.
Further, not actually sure this is nice idea
to use KLT in such synthetic case.
commit 1e22cbcac480863b6b5abc5c85f23dc70748933a
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Sat Apr 6 16:40:59 2013 +0600
Pyramid tracker unit test fix
Issue was caused by trackers modifying guessed
point location even in case of failure. So made
id so both level 0 and level 3 of pyramid are
starting from the same initial guessed location.
Modifying locations in case tracker returns false
is not actually a bug - someone could still want
to use that location. False in this case means
more like "returned location is not so much
accurate".
commit 3ad5e0efa071f202ee7c2034d70dd97aa62b13aa
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Sat Apr 6 00:38:40 2013 +0600
Intersect unit test fix
EuclideanIntersect is not aware of camera calibration
matrix yet and always assumes it to be an identity.
So using non-identity matrix to construct sample case
leads to wrong projection results.
For now made it so test case uses identity matrix for
calibration.
Also fixed variable shadowing which lead to wrong
markers positions (were either zero or undefined).
commit 3a153c2b65f38653a36c36975018f68d42d60670
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Sat Apr 6 00:12:12 2013 +0600
Camera intrinsics unit tests fixes
- Existing test ApplyIsInvertibleSimple was not
doing right thing - distortion model used there
was ininvertible.
Tweaked parameters in a way model is invertible now.
- Added some own tests which tests;
* Principal point always maps from pixel space to
zero normalized position.
* Some basic tests to check whether individual
apply/invert works correct.
commit e3b2bccba6145290738a6677c14f7369ec7a38cd
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Thu Apr 4 02:59:58 2013 +0600
Suppress strict compiler warnings in glags/glog libraries
commit 5fca459adcf0a3419fa9cd8d983dc2c952d02647
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Thu Apr 4 01:20:18 2013 +0600
Lint cleanup, mostly white space and line width.
Also moved own includes to the top of files.
Should be no functional changes :)
commit 9a9dd458a622928b91dbd3c79900577923283838
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Fri Mar 29 00:20:29 2013 +0600
Fix for TransformTracks in uncalibrated pipeline
Transformation matrix was completely ignored by
TransformTracks() and final marker coordinate
exactly matched it's source coordinates.
Seems to be just a typo in vector usage: need to
use "b" (which is transformed one) instead of "a"
when converting projective coordinate to 2D space.
commit d35766cc9901609e32f4d80faba715695bea3c40
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Fri Mar 29 00:19:11 2013 +0600
Revert part of e2eb58c4230f94ef0c72fb4005e5434088d52e80
That commit included one change which shall have been
go as separate commit with more detailed description.
commit e8d71b4e96fd78eb60773b6557d66da672e65753
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Wed Mar 27 20:37:05 2013 +0600
Silenced more warnings
- Added includes of own header to fast implementation files.
- Camera intrinsics wouldn't complain about unknown pragma when
building without OpenMP support.
TODO: Make it a CMake option to build libmv with OpenMP support.
Currently multi-threaded intrinsics only available when
using custom CMake rules for bundled libmv version
(as it's done in Blender).
commit ad442812654f270dc088394410fda1b81b8dc450
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Wed Mar 27 20:18:51 2013 +0600
Multithreaded camera intrinsics
Implemented multithreaded buffer (un)distortion
for camera intrinsics using OpenMP.
By default, (un)distortion is single-threaded,
but it is possible to as CameraIntrinsics to
use more threads by calling SetThreads method.
commit c88b4881096174a16a9f9e6fc2c9dcad3e255b25
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Wed Mar 27 18:45:09 2013 +0600
Movie functions implementation from panography header
into own CC implementation file.
Before this all panography functions were declared as
static, which is not so much useful from re-useability
point of view.
commit 2d2faf9104bc035722cff6775e1b8e7c93143aba
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Wed Mar 27 18:37:36 2013 +0600
Build shared Ceres library only if BUILD_SHARED_LIBS is enabled
commit daa3ddd3260ccaf2bf9c72eadb89213d91e549ec
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Wed Mar 27 18:21:52 2013 +0600
Update Ceres to upstream version 1.5.0
commit cf5dc678878345ea3f221ce50cb2b9e539c2ab38
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Wed Mar 27 15:06:24 2013 +0600
Code cleanup: removed more deprecated FFmpeg API usage
This time in qt-tracker application.
commit e2eb58c4230f94ef0c72fb4005e5434088d52e80
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Tue Mar 26 17:19:51 2013 +0600
Code cleanup: silent unused variables warnings
commit af89bb24667e39b7e655173ea807fdcfbeef4422
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Tue Mar 26 16:54:14 2013 +0600
Code cleanup: no need to declare empty body for ProgressUpdateCallback:invoke
Make force this method to be overridden by derivative classes.
Also removed currently unneeded callbacks.cc.
commit 0441d4ee06fad0219256a5704f931eec916a3868
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Tue Mar 26 16:37:27 2013 +0600
Code cleanup: silent type narrowing in qt-tracker
commit cd4b61c976448d0fdedefb3ed4b21d70e078f94b
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Tue Mar 26 16:26:39 2013 +0600
Changes to unit testing
- Move ceres test binaries to ${LIBMV_TESTS_OUTPUT_DIR}/ceres,
so they don't mess with libmv's application binaries and
tests.
- Removed ceres_ prefix from ceres unit tests, only use this
prefix for targets (targets need to be unique name).
- Added unit tests data for ceres, otherwise system_test fails.
- Restored "test" makefile target.
commit cf704ada08acc8b26167e7bfb3e1e88fd278de23
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Tue Mar 26 15:01:15 2013 +0600
Code cleanup: use rw-rw-r-- mode for source files
commit 64b31e3e43acb52aaf6f591b9d1c2449bf6ef3bd
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Tue Mar 26 14:57:46 2013 +0600
Code cleanup: don't use deprecated FFmpeg API functions
commit 2a3676499548ad5dba5a5c5eadf3bb71e640b612
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Tue Mar 5 17:40:52 2013 +0600
Switch from DENSE_NORMAL_CHOLESKY to DENSE_QR
DENSE_QR is better behaved numerically and after recent
changes from Sameer there's no big difference in speed.
commit bcb920df02133da5b7e55fbc74edb9222004eecc
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Tue Mar 5 17:15:43 2013 +0600
Update Ceres to 1.5RC3
It brings optimization of DENSE_QR and DENSE_SCHUR solvers.
commit 473996468a4e67e7c860169181a4ff31ce9b8c80
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Fri Mar 1 17:44:54 2013 +0600
Fixed incorrect order of arguments passing
to EXPECT_EQ in keyframe selection tests.
commit d38ebb74693fdf5b8f0fecf62a3d8c9c53b0b84a
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Fri Mar 1 17:40:38 2013 +0600
Modal (aka tripod) solver rework
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!
commit 5d6c2e7a27bdd1a1b23bf289d70a9b8f62514c9a
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Fri Mar 1 17:37:35 2013 +0600
Increase verbosity level for reprojected markers info
This information is useful, but in cases when you, say,
working on a bundler it's annoying to scroll all the
information up.
commit ac252bb1250b3028b9c94736b644e7ab4e7b14b8
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Fri Mar 1 17:36:19 2013 +0600
Move radial distortion code to own templated function
This shall not lead to any functional changes, just
avoids radial distortion code duplicated in camera
intrinsics and bundling code.
For fancier bundle adjustment support of 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.
commit c253b794612dd529e1d3a9bd7a7c41c32c9a9abb
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Fri Mar 1 17:33:27 2013 +0600
Use threaded cost function, jacobian and linear solver
computation, so bundling is as fast as it could be with
current parameter block structure.
commit 931fe37a10212b91b525d4f6eb753990a338b471
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Fri Mar 1 17:29:21 2013 +0600
Fixed comment for euclidean bundling,
which is now supports raidal bundling independently
from other intrinsics.
commit 217d8e6edc3de1a853fb84275d2d2dd898e7529c
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Tue Feb 26 18:19:01 2013 +0600
Allow K1,K2 refirement combination
It is now possible to refine only radial distortion
with new Ceres based bundler and this new combination
is already used in Blender.
commit d8850addc944d400f7a9c358396c437d9e4acc70
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Tue Feb 26 18:17:09 2013 +0600
Switch euclidean intersection code to use Ceres
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.
commit 6990b7946ec96b3cb2dcfc8a1beaaba9538b0802
Author: Keir Mierle <mierle@gmail.com>
Date: Mon Feb 25 20:00:48 2013 +0000
Switch motion tracker bundle adjustment to Ceres.
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.
commit 1696342954614b54133780d74d6ee0fbcbe224f0
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Tue Feb 26 18:10:33 2013 +0600
Upgrade ceres to latest upstream version
This is needed because of some features of new Ceres
for further development.
commit 575336f794841ada90aacd783285014081b8318c
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Mon Jan 7 15:58:40 2013 +0600
Fixed for keyframe selection
- Calculate residuals for GRIC in pixel space rather than
in normalized space.
This seems to be how it's intended to be used.
Algebraic H and F will still use normalized coordinates which
are more stable, after this matrices are converted to pixel
space and Ceres refinement happens in pixel space.
- Standard deviation calculation was wrong in GRIC. It shouldn't
be deviation of residuals, but as per Torr it should be deviation
of measurement error, which is constant (in our case 0.1)
Not sure if using squared cost function is correct for GRIC,
but cost function is indeed squared and in most papers cost
function is used for GRIC. After some further tests we could
switch GRIC residuals to non-squared distance.
- Bring back rho part of GRIC, in unit tests it doesn't make
sense whether it's enabled or not, lets see how it'll behave
in real-life footage.
- Added one more unit test based on elevator scene and manual
keyframe selection.
commit 24117f3c3fc5531beb6497d79bb6f1780a998081
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Sun Jan 6 19:07:06 2013 +0600
Added test for keyframe selection based on manual selection
Additional changes:
- Reduce minimal correspondence to match real-world manually
tracked footage
- Returned back squares to SymmetricEpipolarDistance and
SymmetricGeometricDistance -- this is actually a cost
functions, not distances and they shall be squared.
commit 770eb0293b881c4c419c587a6cdb062c47ab6e44
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Fri Dec 21 00:43:30 2012 +0600
Improvements for keyframe selection
- Changed main keyframe selection cycle, so in cases there're no
more next keyframes for current keyframe could be found in the
image sequence, current keyframe would be moved forward and
search continues.
This helps in cases when there's poor motion in the beginning
of the sequence, then markers becomes occluded. There could be
good keyframes in the middle of the shot still.
- Extended keyframe_selection_test with real world cases.
- Moved correspondences constraint to the top, so no H/F estimation
happens if there's bad correspondence. Makes algorithm go a bit
faster.
Strangely, but using non-squared distances makes neighbor frames
test fail, using squared distances makes this tests pass.
However, using non-squared distances seems to be working better
in real tests i've been doing. So this requires more investigation/
commit 7415c62fbda36c5bd1c291bc94d535a66da896d0
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Thu Dec 20 18:46:09 2012 +0600
Cosmetic change to correspondences reports in keyframe selection
commit ceaf80c987ec0338e7e83965bc808411453eb755
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Thu Dec 20 18:08:03 2012 +0600
Various fixes:
- That was a typo in symmetric geometric cost functor, which
computed inverse distance in a wrong way.
- Fixed compilation of unit tests
- Added simple test for keyframe selection. Currently only
covers case that neighbor frames with only translation
(homography should be better than fundamental) are not
considered a keyframes.
Still need to be investigated why it only works if tracks
are in pixel space and why doesn't work in normalized space.
commit cfabdfe48df2add3d1f30cf4370efd0b31990ab0
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Thu Dec 20 05:46:53 2012 +0600
Assorted fixes for keyframe selection:
- Biggest error was in cost functors used for F and H refirement,
they were just wrong.
- Use natural logarithms, since it's actually makes sense from
math papers point of view and error is somewhere else.
- Disabled rho for GRIC, for now use non-clamped error for tests.
- Made SymmetricEpipolarDistance returning non-squared distance
Keyframe selection is currently the only used of this function
and it seems using non-squared distance makes much more sense.
Also would think to append suffix "Squared" to functions which
returns squared distances.
- Removed templated version of SymmetricEpipolarDistance, since
it's not needed actually.
This is actually even worse working than previous implementation,
but commit it needed for further review.
commit 35d8c57626ad74818f155e6e5960c663ea84e032
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Thu Dec 20 03:00:40 2012 +0600
Euclidean resection cost function didn't use correct constructor
It was storing a reference to initial rotation passed by value,
leading to pointer being pointing to a stack variable, leading to
wrong memory access in residuals computing.
Apparently was visible in optimized builds only with inline
substitution allowed.
commit 0798d3162bb49cee7e1c423ceccbca1326ad5650
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Thu Dec 20 02:50:52 2012 +0600
Automatic keyframe selection based on Pollefeys's criteria
This commit implements automatic keyframe selection algorithm
based on Pollefeys's criteria (F-GRIC is smaller than H-GRIC
and correspondence ratio is more then 90%).
It is implemented as a part of simple pipeline and returns
vector of keyframe images for a given Tracks structure.
For simple pipeline reconstruction two best keyframes are
expected to be selected from all detected candidates.
Criteria for this selection could be reprojection error of
solution from two candidate keyfames.
Unfortunately, it's not fully workable yet, hopefully would
be fixed soon.
commit e943985552f0598ae122252876f305d72c25c2f9
Author: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Thu Dec 6 17:47:11 2012 +0600
Camera Tracking: allow fallback to reprojection resection
by user demand
This fixes some "regressions" introduced in previous commit
which lead to much worse solution in some cases. Now it's
possible to bring old behavior back.
Perhaps it's more like temporal solution for time being smarter
solution is found. But finding such a solution isn't so fast,
so let's bring manual control over reprojection usage.
But anyway, imo it's now nice to have a structure which could
be used to pass different settings to the solver.