cashiwamochi / simplebundleadjustment

C++ implementation for Bundle Adjustment in 2-View

Bundle Adjustment 2-View

This repository includes a C++ implementation which performs Bundle Adjustment in 2-View. This BA is very simple, so it maybe a tutorial for beginners. This supports pose-only-BA, structure-only-BA, full-BA. These reference pdf are very helpful.

[1] http://frc.ri.cmu.edu/~kaess/vslam_cvpr14/media/VSLAM-Tutorial-CVPR14-A13-BundleAdjustment.pdf   
[2] https://pdfs.semanticscholar.org/6081/417b95bec070fb842a704044def427f8ef69.pdf   
[3] http://ingmec.ual.es/~jlblanco/papers/jlblanco2010geometry3D_techrep.pdf   
[Japanese]
[4] https://ipsj.ixsq.nii.ac.jp/ej/?action=repository_action_common_download&item_id=62864&item_no=1&attribute_id=1&file_no=1

Implementation

Using Gauss-Newton method, the projection-error is minimized.

BundleAdjustment2Viewes.* -> Implementation of BA.
example_ba.cc -> An example shows how to use BA.

Dependencies

OpenCV > 3.0
OpenCV is used for initialization and calculation of matrices.

Build & Use

mkdir build && cd build && cmake .. && make
./ba_example [path/to/image1] [path/to/image2]

---

2-Viewでのバンドル調整

2-Viewにおけるバンドル調整を実装した．pose-only-BA, structure-only-BA, full-BA を一通りサポートしている．実装にあたっては下記資料が大変参考になった．

[1] http://frc.ri.cmu.edu/~kaess/vslam_cvpr14/media/VSLAM-Tutorial-CVPR14-A13-BundleAdjustment.pdf   
[2] https://pdfs.semanticscholar.org/6081/417b95bec070fb842a704044def427f8ef69.pdf   
[3] http://ingmec.ual.es/~jlblanco/papers/jlblanco2010geometry3D_techrep.pdf   
[日本語]
[4] https://ipsj.ixsq.nii.ac.jp/ej/?action=repository_action_common_download&item_id=62864&item_no=1&attribute_id=1&file_no=1

実装

まず，2枚の画像で特徴点を対応づけ，カメラ姿勢の推定，そして三角測量を行う．こうして得られたカメラ姿勢，三次元点のどちらか，あるいは両方にノイズを加え意図的に再投影誤差を発生させる(なお，ノイズを与えなくても持っている初期値には改善の余地があるため多少誤差を下げられる)．その後，更新したい成分についてヤコビアンを計算し，ガウスニュートン法によって再投影誤差を最小化するという一般的な流れ．カメラ姿勢や三次元点のパラメタライズには特に "A tutorial on SE(3) transformation parameterizations and on-manifold optimization" [3] を参考にした．

BundleAdjustment2Viewes.* -> バンドル調整の実装
ba_example.cc -> 2画像の特徴点対応づけ，カメラ姿勢推定，三角測量を行い，その後バンドル調整を行う．

依存

OpenCV > 3.0
OpenCVは初期化と行列計算のために必要．

ビルド

mkdir build && cmake .. && make
./ba_example [path/to/image1] [path/to/image2]

simplebundleadjustment's Issues

Two different intrinsic matrices for the 2-View

How would the Jacobian ComputeJ change when we have two different intrinsic matrices.

Why adding noise to target data for bundle adjustment?

Why is the target data added with noise of the initial estimate?

SimpleBundleAdjustment/ba_example.cc

Line 183 in 2b4e524

optimizer.SetTargetData(std::vector<cv::Mat>{Rt1_noise, Rt2_noise});

When I add noise, I see that the bundle adjustment is converging at my initial estimate itself. So why add the noise?

About delta pose update

it seems that you compute delta pose on se(3) with left disturbance model, but I noticed that you update the pose with right multiply? why?

‘’‘
// Update!
vm_poses[i] = vm_poses[i] * delta_SE3;
‘’’

[WARN] : The pair of images is empty, it is need in debug.

Hello, may I ask, the program input only has 2 pictures, why can I get 2 poses, and what does the warning mean?

One question is, why in FULL optimization, the dimensionality of the Jacobian matrix has to be 5 degrees of freedom?
Another question is, why can't g2o be optimized for only 3D points, but the program can do it here?
Tks very much. Hope helping.