Global SfM problem

[MEGA728]

I used IncrementalSfM with pre-calculated correspondences , and it worked fine. But when I used GlobalSfM, this happened "-- #Relative triplet of translations estimates: 0 computed from 969 triplets."

Can anybody tell me why it became 0 ?

Thanks!

[pmoulon]

Hi @MEGA728 Can you share the reconstruction log?

The cause could be due to some data intrinsic that are too bad or due to the fact that the computed global rotations are false.

[MEGA728]

Hi @pmoulon

Thank you for your fast reply !

What do you mean by "reconstruction log"? I found this "Reconstruction_Report.html" and attached it.

Another question: after calibrating my camera, I got two focal length, fx and fy, but in sfm_data.json there is only one "focal" in the intrinsics, how do I use the pre-calibrated fx and fy?

[pmoulon]

I mean the log that is shown on the console output.

You can use f = (fx+fy)/2 +> must be in pixels Be sure to configure also the principal point position

[MEGA728]

Ah, the log is as follows: GlobalSfM.txt

[pmoulon]

From the log I saw that the rotation averaging for your dataset produce a large residual error for the triplet composition error => median: 17.3482

Often the value is below 5 or 2 degree. Do not know what is happening without playing with your data.

[MEGA728]

Some days ago, I used Global SfM to deal with another pre-calculated correspondences dataset, residual error for the triplet composition error => median: 15.4145, which is also quite large, but I can still get a good result from Global SfM.

About the current problem, the data is as follows, you can take a look at it, if it is not too much trouble :) input.zip

I was just thinking that maybe this problem occurred because I used too few matches? I only used 15 points from the target shown bellow .

[pmoulon]

@MEGA728 I made a fast test on your dataset, and I was able to get a nice result with the Incremental pipeline. For the globalSfM pipeline the default settings ask more points per pair than that you have (more than 15).

The incremental pipeline seems to provide a very nice result (according the median residuals (0.0245262) and the produced point cloud)

-------------------------------
-- Structure from Motion (statistics):
-- #Camera calibrated: 19 from 19 input images.
-- #Tracks, #3D points: 15
-------------------------------

SequentialSfMReconstructionEngine::ComputeResidualsMSE.
    -- #Tracks: 15
    -- Residual min:    9.45625e-05
    -- Residual median: 0.0245262
    -- Residual max:     0.219888
    -- Residual mean:    0.032899

./openMVG_main_IncrementalSfM -i /home/pierre/Pictures/Datasets/mega72
8/input/sfm_data.json -m /home/pierre/Pictures/Datasets/mega728/input/ -o /home/pierre/Pictures/Datasets/mega728/output_inc

Here as ortho view

For the GlobalSfM failure (15 degree mean error for cycle relative rotation composition) => Seems like your point configuration is perfectly planar, it looks like a degenerate configuration that I think make some relative essential matrix decomposition fail. Perhaps a Homography fit plus homography decomposition would provide a better results.

[MEGA728]

I get it, the GlobalSfM may not be suitable for planar object. THX !

[pmoulon]

It may if we have the right initial Relative rotation decomposition for planar object ;-)

[MEGA728]

So sometimes we get the the right initial relative rotation decomposition, sometimes we don't. Is there any method to assure that we have the right initial relative rotation decomposition for planar object every time?

[pmoulon]

Using Homography and Homography (R|t) decomposition.

[MEGA728]

Thanks for your advice!