Hi everybody, I'm currently trying to incorporate Matlab's bundleAdjustment function into my project. While it seems to work in the examples (even though there is no ground thruth to compare the results), in my project the refined camera poses and 3d points are far from the true solution.
I am evaluating the motion of a camera moving exactly in the -x (left) direction. The images are captured exactly 50mm apart. A point cloud is triangulated in the first two frames and the true scale is given knowing the length of the motion. From the 3rd frame, p3p-RANSAC is used to retrieve the global camera positions. Let's say I compute 5 camera poses and i compute also the reprojection error of the triangulated cloud. The mean reprojection error is approx 1 pixel for each point. The reconstruction of these 5 views is pretty accurate, the maximum error is approximately 5mm.
Now I want to refine those 5 poses using the bundle adjustment function embedded in Matlab and here I have issues.
->If I use the default settings the function will not Iterate more than once, saying that the variation of the reprojection error is below the threshold
->If I set ... 'AbsoluteTolerance', 1e-9, 'RelativeTolerance', 1e-9, 'MaxIterations', 300 (as in the Monocular Visual Odometry example) and fix the two initial camera position, the optimization will converge to wrong results.
Now I'm asking if someone that has succesfully implemented this function have experienced some issues or if there are some other factors to consider while performing bundle adjustment.
Thank you
