OpenCV Bundle Adjustment

Question

I am trying to use the OpenCV Bundle Adjustment: LevMarqSparse::bundleAdjust Unfortunately the output seems totally wrong:

Iteration: 0, normError: 4.30244e+46 (3.64614e+44)
decreasing lambda to 0.024556
Iteration: 1, normError: 2.72985e+31 (2.31343e+29)
decreasing lambda to 0.0024556
Iteration: 2, normError: 8.00302e+25 (6.78222e+23)
move failed: lambda = 0.024556, e2 = 2.62043e+40 (2.2207e+38) > 8.00302e+25(6.78222e+23)
move failed: lambda = 0.24556, e2 = 7.31034e+40 (6.1952e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556, e2 = 1.59706e+40 (1.35344e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 24.556, e2 = 1.33882e+40 (1.13459e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 245.56, e2 = 1.31518e+40 (1.11456e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2455.6, e2 = 1.31283e+40 (1.11257e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 24556, e2 = 1.3126e+40 (1.11237e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 245560, e2 = 1.31257e+40 (1.11235e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556e+06, e2 = 1.31257e+40 (1.11235e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556e+07, e2 = 1.31257e+40 (1.11235e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556e+08, e2 = 1.31257e+40 (1.11235e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556e+09, e2 = 1.31257e+40 (1.11235e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556e+10, e2 = 1.3125e+40 (1.11229e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556e+11, e2 = 1.31102e+40 (1.11103e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556e+12, e2 = 1.3103e+40 (1.11042e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556e+13, e2 = 1.31471e+40 (1.11416e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556e+14, e2 = 1.29156e+40 (1.09454e+38) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556e+15, e2 = 1.15709e+40 (9.80585e+37) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556e+16, e2 = 3.81423e+39 (3.2324e+37) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556e+17, e2 = 6.59182e+36 (5.58629e+34) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556e+18, e2 = 8.07694e+34 (6.84487e+32) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556e+19, e2 = 7.73149e+52 (6.55211e+50) > 8.00302e+25 (6.78222e+23)
move failed: lambda = 2.4556e+20, e2 = 2.061e+40 (1.74661e+38) > 8.00302e+25 (6.78222e+23)
decreasing lambda to 2.4556e+19

If I manually reproject my first guesses for my keypoints I get a total(summed) euclidean distance in pixels(error) of 169.541 between each keypoint and the reprojection of the it's triangulated point. The same error-check after the call to LevMarqSparse::bundleAdjust gives an error of 5.50025e+12.

Could someone please point me to the right direction? Thanks.

Answer 1

In general, when something like this happens with bundle adjustment at least one of the following errors are very likely:

1. The 2D projections are not the correct ones for the 3D point, so that one 3D point is reprojected onto false 2D projections and thus, the reprojection error will be very large. Perhaps something has happened with the indices between the 2D projections and the 3D points.
2. The cameras you use are not in the same coordinate system like OpenCV requires them to be. Imagine a rotation about 180° or an inverse movement and the 3D points will be projected somewhere but surely not on the correct 2D locations. So, perhaps try to take the inverse camera matrices. If P is one of your 4x4 homogeneous projection matrices take P.inv() instead of P.

It would help to get the mean reprojection error. The sum is not very meaningful since we don't know how many 3D points and 2D projections you do have in your reconstruction.