Pleasure to see the wonderful work! I read the paper and did not find any words to explain how to exploit the x^beta. In other words, how to fuse the results of common parameters from respective k-th iteration of neighboring devices? Any explanation will be appreciated!

If I add pose graph error and jacobians, it can solve also?

I don't have the knowledge about mpi / nccl. so I don't know about it works well.

I want to know DABA solve pose graph problems, when I use your api about Least square optimization.

Thank you.

I think it's possible in principle, but maybe very very hard to implement technically.

Hi,

First, thanks for your great work.

Because I'm not sure what tool to use to visualize the results, so I transform the "results-mm-lm-acc-trivial-problem-1723-156502-pre-2-GPU-1000-iters.txt" to the colmap format(which save r_{iw}, t_{iw}, p_w).

In my understand, "results-mm-....txt" saves:
[R_{wi} | R_{wi}t_{iw} = -t_{wi}]
f k1 k2
-p_w1 -p_w2 -p_w3

After I have transformed the poses and points, I got a weird result.

(after optimization)
this is the code I try to transform

  extrinsic.resize(num_extrinsics);
  for (int i = 0; i < num_extrinsics; i++) {
    std::getline(file, line);
    Eigen::Matrix<double, 3, 3> r;
    Eigen::Vector3d t;
    for (int m = 0; m < 3; m++) {
      std::getline(file, line);
      std::istringstream iss(line);
      iss >> r.coeffRef(m, 0) >> r.coeffRef(m, 1) >> r.coeffRef(m, 2) >> t[m];
    }
    // if r is r_wi
    // t_iw = - r_wi^T * t_iw =r_wi^T  *( - t_wi) = r_wi^T * t
    Eigen::Vector3d r_aa;
    Log(r.transpose(), r_aa); // copy from DABA/sfm/math/
    t = r * t;
    Eigen::AngleAxisd r_aad(r_aa.norm(), r_aa.normalized());
    Eigen::Quaterniond qua(r_aad);
    qua.normalize();
    qua.inverse();

    extrinsic[i].resize(7);
    extrinsic[i][0] = qua.w();
    extrinsic[i][1] = qua.x();
    extrinsic[i][2] = qua.y();
    extrinsic[i][3] = qua.z();

    for (int m = 0; m < 3; m++) {
      extrinsic[i][m + 4] = t[m];
    }
  }

I transform the input "problem-1723-156502-pre.txt", and I got the correct result:

I did modified you code, and i just run commandmpiexec -n 2 ./bin/mpi_daba_bal_dataset --dataset /path/to/your/dataset --iters 1000 --loss "trivial" --accelerated true --save true.

I don't know where I got a misunderstanding. And I also want to know if there is a tool to visualiza the results easily.

Thank you!

Thank you for your wonderful work, I'd love to look into the details of your project, if you could provide a simple flowchart that would be great!😊

Hi,

First, I'd like to say great work and amazing results.

I've been trying to run DABA on custom-made BAL dataset examples, resulting in failed solutions. Unstable optimization (especially for ceres solver) and divergent final reconstruction. However, standard ceres optimization with Snavely reprojection cost function results in converged output. Additionally, both Snavely ceres implementation, DABA, and your ceres implementation result in the same converged solution when distortion parameters are removed (from both cost and jacobian evaluation) keeping them at zero (basically, keeping simple pinhole camera model).

With that in mind, I have trouble understanding a few parts of the code regarding distortion:

1. In reprojection error cost implementation, the term for undistorted reprojection ray is

, while the derived formula (supported by your paper) is

I'm noticing a slight difference between them in second and fourth order terms.
Even if focal length and measurements were rescaled so that focal length equals 1.0 (in loading methods), second and fourth order terms should still contain the scaled focal length as it is optimized. I'm wondering if there is something I might be overlooking or if there's an explanation for this difference.

2. As I understand, you are using Fitzgibbon's division model for distortion (in pixel space), while BAL/Bundler format uses the polynomial model in camera space. I'm curious to know how these models are interchangeable. I can see that k1 and k2 are scaled with focal length to accommodate between pixel <-> camera space. Can you please elaborate on how this interchangeability is achieved?

My guess on why custom BAL examples are failing, for now, is that the initial problems have large initial errors due to differences in distortion models and that results in divergent behavior. This is supported by the fact that DABA converges if distortion is optimized, but with 0 as the initial value.
If I have misunderstood something, I would appreciate any clarification and hope this helps in running DABA on any bundle adjustment problem.

Thank you!