Check failed: size == existing_size Core dumped #434

riematrix · 2018-07-06T06:03:52Z

Core dumped at structure from motion step.
BUG

Problem

[13:57:31.642496][info] AutomaticInitialPairChoice, test I: 2130908429, J: 2146485851, nbCommonTracks: 308
*
[13:57:31.643868][info] AutomaticInitialPairChoice, test I: 2140548709, J: 2146485851, nbCommonTracks: 140
[13:57:31.682983][info] Initial pair is:
A - view id: 77159844 - filepath: /home/zj/data/szyl_segment/images/DJI_0158.JPG
B - view id: 793985669 - filepath: /home/zj/data/szyl_segment/images/DJI_0160.JPG
[13:57:31.683931][info] 802 matches in the image pair for the initial pose estimation.
problem_impl.cc:101 Check failed: size == existing_size Tried adding a parameter block with the same double pointer, 0x165d370, twice, but with different block sizes. Original size was 7 but new size is 3
Aborted (core dumped)

Conditions

Prepare data set with more than 100 images. Problem may not occur if the amount of images is small.
I've tried with amount 3, 29 and 120, only the 120 dataset failed with image size 5472x3648.

Steps to Reproduce

PYTHONPATH=$PWD python meshroom/ui
Compute StructureFromMotion

Versions

AliceVision branch/version: develop head
OS: ubuntu14.04
C++ compiler: gcc

riematrix · 2018-07-30T04:11:01Z

After some code analysis I added two logs in BundleAdjustmentCeres.cpp at about line 283 and line 404.
Seems that there is a mistake of using ceres-solver, just as the error log shows.

283
double * parameter_block = &map_intrinsics[idIntrinsics][0];
problem.AddParameterBlock(parameter_block, map_intrinsics[idIntrinsics].size());

ALICEVISION_LOG_DEBUG("--------------------AddParameterBlock ");
ALICEVISION_LOG_DEBUG( parameter_block );
ALICEVISION_LOG_DEBUG(" Size ");
ALICEVISION_LOG_DEBUG( map_intrinsics[idIntrinsics].size());
ALICEVISION_LOG_DEBUG( " & " );
ALICEVISION_LOG_DEBUG(&parameter_block);
ALICEVISION_LOG_DEBUG("--------------------" );

404
ceres::CostFunction* costFunction = createCostFunctionFromIntrinsics(sfm_data.intrinsics[view->getIntrinsicId()].get(), observationIt.second.x);

    ALICEVISION_LOG_DEBUG("--------------------start AddResidualBlock int " );
    ALICEVISION_LOG_DEBUG( &map_intrinsics[view->getIntrinsicId()][0] );
    ALICEVISION_LOG_DEBUG( " pose ");
    ALICEVISION_LOG_DEBUG( &map_poses[view->getPoseId()][0] );
    ALICEVISION_LOG_DEBUG( " x " );
    ALICEVISION_LOG_DEBUG( landmarkIt.second.X.data());
    ALICEVISION_LOG_DEBUG("--------------------" );

Here is the out put:

[11:51:58.207050][debug] 518993753, 1782016916 |        +27452.1 |         +1218.0 |           +22.5 |            1214
[11:51:58.207238][info] Initial pair is:
  A - view id: 319589866 - filepath: /home/zj/data/szyl_segment/images/DJI_0181.JPG
  B - view id: 1845795821 - filepath: /home/zj/data/szyl_segment/images/DJI_0183.JPG
[11:51:58.215213][info] 853 matches in the image pair for the initial pose estimation.
[11:52:11.374769][debug] A-Contrario initial pair residual: 52.6677
[11:52:11.647555][debug] BundleAdjustmentCeres: DENSE_SCHUR
[11:52:11.647591][debug] Global BundleAdjustment dense
[11:52:11.647603][debug] BundleAdjustmentCeres: DENSE_SCHUR
[11:52:11.647908][debug] --------------------AddParameterBlock 
[11:52:11.647928][debug] 0x627b950
[11:52:11.647942][debug]  Size 
[11:52:11.647951][debug] 7
[11:52:11.647960][debug]  & 
[11:52:11.647970][debug] 0x7ffe014a7030
[11:52:11.647980][debug] --------------------
[11:52:11.648010][debug] --------------------start AddResidualBlock int 
[11:52:11.648021][debug] 0x627b950
[11:52:11.648032][debug]  pose 
[11:52:11.648040][debug] 0x627bc00
[11:52:11.648049][debug]  x 
[11:52:11.648057][debug] 0x7fe4a4003bc8
[11:52:11.648066][debug] --------------------
problem_impl.cc:101 Check failed: size == existing_size Tried adding a parameter block with the same double pointer, 0x627b950, twice, but with different block sizes. Original size was 7 but new size is 3
The program has unexpectedly finished.
/home/zj/exp/photogrammetry/build-AliceVision-Desktop-Default/Linux-x86_64/aliceVision_incrementalSfM crashed

riematrix · 2018-07-31T03:23:36Z

Some more info about the gdb stack trace and the code analytics.

#0 0x00007f010397cc37 in __GI_raise (sig=sig@entry=6) at ../nptl/sysdeps/unix/sysv/linux/raise.c:56
#1 0x00007f0103980028 in __GI_abort () at abort.c:89
#2 0x00007f010597f415 in MessageLogger::~MessageLogger() () from /home/zj/exp/photogrammetry/AliceVision/build/external/aliceVision_build/lib/libaliceVision_sfm.so.0
#3 0x00007f0105a319c6 in ceres::internal::ProblemImpl::InternalAddParameterBlock(double*, int) () from /home/zj/exp/photogrammetry/AliceVision/build/external/aliceVision_build/lib/libaliceVision_sfm.so.0
#4 0x00007f0105a3225c in ceres::internal::ProblemImpl::AddResidualBlock(ceres::CostFunction*, ceres::LossFunction*, std::vector<double*, std::allocator<double*> > const&) ()
from /home/zj/exp/photogrammetry/AliceVision/build/external/aliceVision_build/lib/libaliceVision_sfm.so.0
#5 0x00007f0105a32c18 in ceres::internal::ProblemImpl::AddResidualBlock(ceres::CostFunction*, ceres::LossFunction*, double*, double*, double*) ()
from /home/zj/exp/photogrammetry/AliceVision/build/external/aliceVision_build/lib/libaliceVision_sfm.so.0
#6 0x00007f010597955b in aliceVision::sfm::BundleAdjustmentCeres::createProblem(aliceVision::sfm::SfMData&, aliceVision::sfm::BA_Refine, ceres::Problem&) ()
from /home/zj/exp/photogrammetry/AliceVision/build/external/aliceVision_build/lib/libaliceVision_sfm.so.0
#7 0x00007f010597b358 in aliceVision::sfm::BundleAdjustmentCeres::Adjust(aliceVision::sfm::SfMData&, aliceVision::sfm::BA_Refine) ()
from /home/zj/exp/photogrammetry/AliceVision/build/external/aliceVision_build/lib/libaliceVision_sfm.so.0
#8 0x00007f010593c56e in aliceVision::sfm::ReconstructionEngine_sequentialSfM::BundleAdjustment(bool) ()
from /home/zj/exp/photogrammetry/AliceVision/build/external/aliceVision_build/lib/libaliceVision_sfm.so.0
#9 0x00007f010594d4a7 in aliceVision::sfm::ReconstructionEngine_sequentialSfM::makeInitialPair3D(std::pair<unsigned int, unsigned int> const&) ()
from /home/zj/exp/photogrammetry/AliceVision/build/external/aliceVision_build/lib/libaliceVision_sfm.so.0
#10 0x00007f010594ea18 in aliceVision::sfm::ReconstructionEngine_sequentialSfM::createInitialReconstruction(std::vector<std::pair<unsigned int, unsigned int>, std::allocator<std::pair<unsigned int, unsigned int> > > const&) () from /home/zj/exp/photogrammetry/AliceVision/build/external/aliceVision_build/lib/libaliceVision_sfm.so.0
#11 0x00007f010595188b in aliceVision::sfm::ReconstructionEngine_sequentialSfM::process() () from /home/zj/exp/photogrammetry/AliceVision/build/external/aliceVision_build/lib/libaliceVision_sfm.so.0
#12 0x0000000000453161 in main ()

Analysis

The error occurs at ceres problem_impl.cc line 101.
This size check happens while calling:

ParameterBlock* ProblemImpl::InternalAddParameterBlock(double* values, int size)

So the size compare is between the size of stored params in parameter_block_map_(existing_size) and the second param (int size) which is passed in.

Trace of those two values:

1. BundleAdjustmentCeres.cpp at about line 282
[11:52:11.647908][debug] --------------------AddParameterBlock

Pointer 0x627b950 which points to one of sfm_data's intrinsic params, was the first added to map_intrinsics.
At the same time this pointer(parameter_block) was added to ceres's Problem at line 283:

problem.AddParameterBlock(parameter_block, map_intrinsics[idIntrinsics].size());

Function AddParameterBlock will call ceres's function InternalAddParameterBlock.
So I guess map_intrinsics[idIntrinsics].size() with value 7 is the first value which is stored in ceres's parameter_block_map_?

2. BundleAdjustmentCeres.cpp at about line 404
[11:52:11.648010][debug] --------------------start AddResidualBlock

  problem.AddResidualBlock(
      costFunction,
      p_LossFunction,
      &map_intrinsics[view->getIntrinsicId()][0],
      &map_poses[view->getPoseId()][0],
      landmarkIt.second.X.data()); //Do we need to copy 3D point to avoid false motion, if failure ?
  }

The third param &map_intrinsics[view->getIntrinsicId()][0] is just the same pointer 0x627b950, which is the former intrinsic params.
The InternalAddParameterBlock was called again at ceres problem_impl.cc line 276:

  InternalAddParameterBlock(parameter_blocks[i],
                              parameter_block_sizes[i]);

So from the error we know the parameter_block_sizes[i] is 3. Trace back the second param parameter_block_sizes[i], it was somethig about
the passed in cost function at line 245:

const vector<int32>& parameter_block_sizes =
  cost_function->parameter_block_sizes();

The cost function was created by BundleAdjustmentCeres.cpp at about line 403:

 ceres::CostFunction* costFunction = createCostFunctionFromIntrinsics(sfm_data.intrinsics[view->getIntrinsicId()].get(), observationIt.second.x);

If the trace is correct, parameter_block_sizes() will return one internal value which is initalized during the definition of the cost function.

I'm sure we are using the PINHOLE_CAMERA modal across all images. So the value 3 is the intrinsics size, i.e. the second number at line 26:

case PINHOLE_CAMERA:
  return new ceres::AutoDiffCostFunction<ResidualErrorFunctor_Pinhole, 2, 3, 6, 3>(
    new ResidualErrorFunctor_Pinhole(observation.data()));

Finally, seems that those two values should not match proberbly because they have different meanning. Perhaps one is the real size of intrinsics and the other is the size of target intrinsics to be optimized? I have no idea which is correct during BundleAdjustment.

The same problem may exist in LocalBundleAdjustmentCeres.cpp. Hope this info will hlep. Thx.

gregoire-dl · 2018-07-31T10:42:55Z

Hi,

Thank you for your detail analysis.
It's a great start for our investigation.
Is it possible to send us your images and meshroom scene file?
(you can do it privately using alicevision-team@googlegroups.com)

riematrix · 2018-08-01T03:51:15Z

Hi,
I've already shared the data via email.

PS
File cameraInit.sfm in MeshroomCache contains "intrinsics": [
{
"intrinsicId": "2851613064",
"width": "5472",
"height": "3648",
"type": "pinhole",
"serialNumber": "c9dbe393a1e134e2cd33ae245d5e1ceb",
"pxInitialFocalLength": "-1",
"pxFocalLength": "8.8000000000000007",
"principalPoint": [
"2736",
"1824"
],
"distortionParams": [
"0",
"0",
"0",
"0"
],
"locked": "false"
}
]

Although distortionParams are zeros, they are add to camera model at Pinhole.hpp line 112

params.insert(params.end(), _distortionParams.begin(), _distortionParams.end());

so the result vector size of getParams() will return 7, which will not match size 3 defined by the pinhole cost function?

fabiencastan · 2018-08-01T08:02:12Z

Do you manually changed the camera model on the CameraInit to Pinhole in Meshroom?

riematrix · 2018-08-01T10:22:08Z

@fabiencastan yes, is it should remains to be unchanged?

fabiencastan · 2018-08-01T10:25:08Z

No, you are correct, it is definitely a bug! There is an old PR to fix it that has never been merged.
I just updated it:
#347
Can you try it?

fabiencastan · 2018-08-01T10:26:52Z

I just tried your dataset and it works for me with the PR.

fabiencastan · 2018-08-01T18:27:24Z

It's now merged in the "develop" branch.

riematrix · 2018-08-02T09:19:46Z

Fixed with #347

simogasp added the type:bug label Jul 30, 2018

riematrix closed this as completed Aug 2, 2018

natowi added the linux label Jun 30, 2019

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Check failed: size == existing_size Core dumped #434

Check failed: size == existing_size Core dumped #434

riematrix commented Jul 6, 2018 •

edited

Loading

riematrix commented Jul 30, 2018 •

edited by simogasp

Loading

riematrix commented Jul 31, 2018 •

edited

Loading

gregoire-dl commented Jul 31, 2018

riematrix commented Aug 1, 2018 •

edited

Loading

fabiencastan commented Aug 1, 2018

riematrix commented Aug 1, 2018 •

edited

Loading

fabiencastan commented Aug 1, 2018

fabiencastan commented Aug 1, 2018

fabiencastan commented Aug 1, 2018

riematrix commented Aug 2, 2018

Check failed: size == existing_size Core dumped #434

Check failed: size == existing_size Core dumped #434

Comments

riematrix commented Jul 6, 2018 • edited Loading

Problem

Conditions

Steps to Reproduce

Versions

riematrix commented Jul 30, 2018 • edited by simogasp Loading

riematrix commented Jul 31, 2018 • edited Loading

Analysis

gregoire-dl commented Jul 31, 2018

riematrix commented Aug 1, 2018 • edited Loading

fabiencastan commented Aug 1, 2018

riematrix commented Aug 1, 2018 • edited Loading

fabiencastan commented Aug 1, 2018

fabiencastan commented Aug 1, 2018

fabiencastan commented Aug 1, 2018

riematrix commented Aug 2, 2018

riematrix commented Jul 6, 2018 •

edited

Loading

riematrix commented Jul 30, 2018 •

edited by simogasp

Loading

riematrix commented Jul 31, 2018 •

edited

Loading

riematrix commented Aug 1, 2018 •

edited

Loading

riematrix commented Aug 1, 2018 •

edited

Loading