rigid icp merged not well

[ShiwenH]

Hi I managed to run the rigid_icp example and get a very good result. However when i tried to import frame_1.ply and frame_2.ply into this demo (by changing some example code), I found that these two files are not merged well. It means this is a not small gab between the transformed data.

Could you help to explain. I do not like pcl because it is so heave and slow.

[kzampog]

Hi!

Those two point clouds were meant as input to non_rigid_icp, since they are not related via a rigid transformation (so this is generally expected). If you suspect that rigid ICP behaves erratically on them, please share some code and sample output, and I will investigate!

Thanks!

[ShiwenH]

Hi,

Today I scanned 350 frames with a 3D sensor and then imported the beginning 50 frames into my demo based on this SDK. The result is not bad. but for the beginning 100 frames, it is too bad.

Here is the sample data. (uploading...will update)

Here is my demo source code. (In this code, I use a new tinyply because I do not know how to use old tinyply you use. you can adjust slightly the saving ply part.) rigid_icp.txt

I have checked that these frames can be registrated well in some other tools.

Regards.

[kzampog]

Hi! I'll have a look once the data is uploaded!

[ShiwenH]

https://www.dropbox.com/s/xbfbtcm009r87s8/model_back.7z?dl=0

This is the dropbox link.

Thanks so much.

[kzampog]

Hi!

I just did some testing with your code and data. I increased the correspondence distance threshold to 100 and noticed that failures then start after frame 90. One failure happens when registering frames 91 and 92: point-to-plane does align the table surfaces but there is not enough geometry to disambiguate for sliding (e.g., more surfaces not parallel to the table would help)!

One solution you could consider is using frame-to-model alignment, which is less brittle (but slower if done naively as below):

#include <cilantro/icp_common_instances.hpp>
#include <cilantro/point_cloud.hpp>
#include <iostream>

int main(int argc, char ** argv) {
    cilantro::RigidTransform3f cumulativeMatrix = cilantro::RigidTransform3f::Identity();
    cilantro::PointCloud3f cumulativeCloud;

    char dstName[100];
    char srcName[100];

    for (int i = 0; i < 352; i++) {
        sprintf(dstName, "/home/kzampog/Desktop/icp_test/model/model_back/1_%05d.ply", i);
        sprintf(srcName, "/home/kzampog/Desktop/icp_test/model/model_back/1_%05d.ply", i + 1);

        cilantro::PointCloud3f dst(dstName), src(srcName);

        if (dst.isEmpty() || src.isEmpty()) {
            std::cout << "One of the clouds is empty!" << std::endl;
            return 0;
        }

//        dst.transform(cumulativeMatrix);
        src.transform(cumulativeMatrix);

//        cilantro::SimpleCombinedMetricRigidICP3f icp(dst.points, dst.normals, src.points);
        cilantro::SimpleCombinedMetricRigidICP3f icp(cumulativeCloud.points, cumulativeCloud.normals, src.points);

        // Parameter setting
        icp.setMaxNumberOfOptimizationStepIterations(1).setPointToPointMetricWeight(0.0f).setPointToPlaneMetricWeight(1.0f);
        icp.correspondenceSearchEngine().setMaxDistance(100.0f);
        icp.setConvergenceTolerance(1e-4f).setMaxNumberOfIterations(30);

        cilantro::RigidTransform3f tf_est = icp.estimate().getTransform();

        std::cout << "Iterations performed: " << icp.getNumberOfPerformedIterations() << std::endl;
        std::cout << "Has converged: " << icp.hasConverged() << std::endl;
        std::cout << "ESTIMATED transformation R:" << std::endl << tf_est.matrix() << std::endl;

        if (i == 0) {
            cumulativeCloud.append(dst);
        }

        src.transform(tf_est);

        cumulativeCloud.append(src).gridDownsample(1.0f);

        cumulativeMatrix = tf_est * cumulativeMatrix;
    }

    cumulativeCloud.toPLYFile("/home/kzampog/Desktop/a1.ply");

    return 0;
}

I also included a downsampling step on model update, but probably not with the best bin size. With the above, all frames in the sequence were registered correctly:

If your data comes from an RGBD camera, you could also build upon the included fusion example (which is roughly like KinectFusion using a point-based model representation) -- that should improve the model quality quite a bit.

Please let me know if I can be of further help.

Cheers!

[ShiwenH]

Wow. Excellent. Last Sunday, I was attending my brother's wedding. Sorry for responding you so late. Tonight I will test it with this model and any other models. Thanks so much. After testing it, I will learn this code in detail. I really like this source code because it is faster and lighter than PCL.

[kzampog]

Hehe no worries -- congrats to your brother! I am really glad you like this :)

[ShiwenH]

Hi. I just tested the code you provided. The result was awesome but the speed was a bit higher because I might use it for cross-platform. When I set the gridDownsample value to 3.0 and the max iteration number to 20, processing 350 frames costed 180 seconds. Are there any suggestions to further accelerate it? Besides, I want to learn this project. It would be great that you provide some relevant papers.

[kzampog]

Hi!

It takes about 30 seconds on my 4C/8T machine (with the values you suggested). Did you compile with OpenMP?

This is the code I tested with (slight modification of the previous version):

#include <cilantro/icp_common_instances.hpp>
#include <cilantro/point_cloud.hpp>
#include <cilantro/timer.hpp>
#include <iostream>

int main(int argc, char ** argv) {
    cilantro::RigidTransform3f cumulativeMatrix = cilantro::RigidTransform3f::Identity();
    cilantro::PointCloud3f cumulativeCloud;

    char pcdName[100];

    cilantro::Timer timer;

    for (int i = 0; i <= 352; i++) {
        sprintf(pcdName, "/home/kzampog/Desktop/icp_test/model/model_back/1_%05d.ply", i);
        cilantro::PointCloud3f pcd(pcdName);

        if (pcd.isEmpty()) {
            std::cout << "Cloud is empty!" << std::endl;
            return 0;
        }

        if (cumulativeCloud.isEmpty()) {
            cumulativeCloud.append(pcd);
            continue;
        }

        timer.start();

        cilantro::SimpleCombinedMetricRigidICP3f icp(cumulativeCloud.points, cumulativeCloud.normals, pcd.points);

        // Parameter setting
        icp.setInitialTransform(cumulativeMatrix);
        icp.setMaxNumberOfOptimizationStepIterations(1).setPointToPointMetricWeight(0.0f).setPointToPlaneMetricWeight(1.0f);
        icp.correspondenceSearchEngine().setMaxDistance(100.0f);
        icp.setConvergenceTolerance(1e-4f).setMaxNumberOfIterations(20);

        cumulativeMatrix = icp.estimate().getTransform();

        timer.stop();
        std::cout << "Frame " << i <<  " registration: " << timer.getElapsedTime() << std::endl;

        // std::cout << "Iterations performed: " << icp.getNumberOfPerformedIterations() << std::endl;
        // std::cout << "Has converged: " << icp.hasConverged() << std::endl;
        // std::cout << "ESTIMATED transformation R:" << std::endl << tf_est.matrix() << std::endl;

        timer.start();

        pcd.transform(cumulativeMatrix);
        cumulativeCloud.append(pcd).gridDownsample(3.0f);

        timer.stop();
        std::cout << "Frame " << i <<  " model update: " << timer.getElapsedTime() << std::endl;

        // cumulativeMatrix = tf_est * cumulativeMatrix;
    }

    cumulativeCloud.toPLYFile("/home/kzampog/Desktop/a1.ply");

    return 0;
}

You could shave off a few seconds by not downsampling at every frame (e.g., do it every 2-5 frames instead), but that's not going to make a big difference. You could also give the fusion example a shot (projective ICP and index map based point fusion); it should be quite a bit faster and produce better quality reconstructions.

Unfortunately, API documentation is sparse at the moment, but the example programs cover a good part of the supported functionality. We also have a technical report on cilantro (link in README.md) that gives a high level overview of the library.

Let me know if I can help with anything!

Cheers!

[ShiwenH]

As visual studio does not support a high enough OpenMP, I tested this demo with VSCode and gcc compiler in windows. After I enabled the OpenMP, the speed is good. However, as we are doing registration with cumulative result, the time will grow faster. Now I only take last 6 frames to do registration and the speed is awesome again.

After migrating from VS to VSCode, the griddownsample speed is very slow. Do you know this issue?

[ShiwenH]

My code attached.

include <cilantro/icp_common_instances.hpp>

include <cilantro/point_cloud.hpp>

include

include

include

int main(int argc, char ** argv) { cilantro::RigidTransform3f cumulativeMatrix = cilantro::RigidTransform3f::Identity();

char firstModel[256];
sprintf(firstModel, "model/model_back/1_00000.ply");
cilantro::PointCloud3f cumulativeCloud(firstModel);

std::deque<cilantro::PointCloud3f> dequeForIcp;
dequeForIcp.push_back(cumulativeCloud);

for (int i = 1 ; i < 42 ; i++)
{
    std::cout << "Index: " << i << std::endl;

    char srcName[256];

    sprintf(srcName, "model/model_back/1_%05d.ply", i + 1);

    cilantro::PointCloud3f src(srcName);

    src.transform(cumulativeMatrix);

    cilantro::PointCloud3f pointcloudForIcp;
    for (int i = 0; i < dequeForIcp.size(); i++)
    {
        pointcloudForIcp.append(dequeForIcp.at(i));
    }

    cilantro::SimpleCombinedMetricRigidICP3f icp(pointcloudForIcp.points, pointcloudForIcp.normals, src.points);

    icp.setMaxNumberOfOptimizationStepIterations(1).setPointToPointMetricWeight(1.0).setPointToPlaneMetricWeight(0.0);
    icp.correspondenceSearchEngine().setMaxDistance(600);
    icp.setConvergenceTolerance(1e-2f).setMaxNumberOfIterations(30);

    cilantro::RigidTransform3f tf_est = icp.estimate().getTransform();

    src.transform(tf_est);

    dequeForIcp.push_back(src);
    if (dequeForIcp.size() > 6)
    {
        dequeForIcp.pop_front();
    }

    cumulativeCloud.append(src);

    cumulativeMatrix = tf_est * cumulativeMatrix;
}

cumulativeCloud.toPLYFile("model/new1.ply");

system("pause");

return 0;

}

[kzampog]

Hi!

That should work well!

How long does gridDownsample take and on what input/parameters? If you could share a test cloud and timing (before/after migrating to gcc), that would be help!

[ShiwenH]

This is the time after adding griddownsample:

This is the new code:

include <cilantro/icp_common_instances.hpp>

include <cilantro/point_cloud.hpp>

include <cilantro/timer.hpp>

include

include

include

int main(int argc, char ** argv) { cilantro::RigidTransform3f cumulativeMatrix = cilantro::RigidTransform3f::Identity();

char firstModel[256];
sprintf(firstModel, "model/model_back/1_00000.ply");
cilantro::PointCloud3f cumulativeCloud(firstModel);

std::deque<cilantro::PointCloud3f> dequeForIcp;
dequeForIcp.push_back(cumulativeCloud);

for (int i = 1 ; i < 76 ; i++)
{
    char srcName[256];

    sprintf(srcName, "model/model_back/1_%05d.ply", i + 1);

    cilantro::PointCloud3f src(srcName);

    src.transform(cumulativeMatrix);

    cilantro::PointCloud3f pointcloudForIcp;
    for (int i = 0; i < dequeForIcp.size(); i++)
    {
        pointcloudForIcp.append(dequeForIcp.at(i));
    }

    cilantro::SimpleCombinedMetricRigidICP3f icp(pointcloudForIcp.points, pointcloudForIcp.normals, src.points);

    icp.setMaxNumberOfOptimizationStepIterations(1).setPointToPointMetricWeight(1.0).setPointToPlaneMetricWeight(0.0);
    icp.correspondenceSearchEngine().setMaxDistance(600);
    icp.setConvergenceTolerance(1e-2f).setMaxNumberOfIterations(30);

    cilantro::RigidTransform3f tf_est = icp.estimate().getTransform();

    src.transform(tf_est);

    dequeForIcp.push_back(src);
    if (dequeForIcp.size() > 5)
    {
        dequeForIcp.pop_front();
    }

    cilantro::Timer timer;
    timer.start();

    cumulativeCloud.append(src).gridDownsample(3.0f);

    timer.stop();

    std::cout << timer.getElapsedTime() << std::endl;

    cumulativeMatrix = tf_est * cumulativeMatrix;
}

cumulativeCloud.toPLYFile("model/new1.ply");

system("pause");

return 0;

}

Regarding the test data, I guess the old data I provided should also work for this test.

[kzampog]

Wow, these numbers are crazy! I am getting about 45ms for the very last downsampling steps (end of sequence), and the whole thing takes about 6.5 seconds on my Ubuntu-based distro. Are you compiling with optimizations enabled (e.g. -O3 for gcc)?

If it's not optimizations, one thing to try would be to disable parallelism for downsampling and see what happens. To do that, try changing line 140 of grid_accumulator.hpp from:

#ifdef ENABLE_NON_DETERMINISTIC_PARALLELISM

to:

#if 0

and see if that changes the times.

[ShiwenH]

-O3 is always there. After #if 0, the time is as followings. It is much better but I guess it is still not perfect.

31.9123 31.9141 37.8982 66.821 54.8529 65.8255 74.7994 78.7888 90.7598 103.722 100.729 107.712 125.665 123.67 138.676 138.629 155.584 158.575 162.594 166.559 183.508 185.503 198.468 192.485 195.477 204.477 219.413 222.414 228.389 238.368 255.366 261.34 257.312 267.286 279.253 290.224 299.241 304.186 319.146 326.166 321.142 346.122 340.09 350.064 353.056 378.014 369.03 371.05 381.978

[kzampog]

So it was parallelism that was slowing it down... In Linux, it does give measurable speedups!

Times now look much more reasonable, but are still a few times higher than what I get on my Linux setup.

I will try to investigate as soon as I get a chance. Thanks for reporting this!

[ShiwenH]

My pleasure