Unable to extract removal ground points

[tuandle]

Hi @niosus , I would like to reproduce your result at Fig3 in your journal paper. I tried a naive approach to get the removal ground points by changing this line from == to =!

if (dilated.at<uint16_t>(r, c) =! 0)

My understanding is that, with this change, depth_ground_remover should return a Cloud obj contains only ground points. However, when visualizing it, I got the original raw pointcloud.

I definitely miss something here and I greatly appreciate your advice.

[niosus]

Ok, this might sound a but counter-intuitive, but for the sake of speed I never updated the points in the cloud after removing the ground from the range image. So the cloud is still the same, but every cloud has a range image attached to it, so the image attached to it is not without ground. After your changes this image will contain ground only, but the cloud does not change.

I set the image here: https://github.com/PRBonn/depth_clustering/blob/57ca59ddf04d40dd761cb550e65c30d497c647e7/src/ground_removal/depth_ground_remover.cpp#L55

[tuandle]

@niosus Thanks for your clarification. I see that everything is based on the range image for fast computation. What I understand so far for removing ground is as follows:

1. You remove the ground based on the range image in depth_ground_remover. The processed range image (should contain non-zero values for non-ground points and zero values for ground point) is sent to image_based_clusterer
2. image_based_clusterer first labels the received range image. It will only label non-ground points (depth pixels with non-zero value). Then it creates a container to store labels and their corresponding points and after that, send the container to the next client.

My approach is like this:

1. I remove all the non-ground points (switch == to !=). The processed range image contains only the ground (with non-zero values; non-ground points are marked as zero) and is sent to the next step.
2. I skip the labeling procedure and create a container which stores only 1 label and all the ground points from the received range image. I create the container like this:

void OnNewObjectReceived(const Cloud& cloud, const int sender_id) override {
    // generate a projection from a point cloud
    if (!cloud.projection_ptr()) {
      fprintf(stderr, "ERROR: projection not initialized in cloud.\n");
      fprintf(stderr, "INFO: cannot label this cloud.\n");
      return;
    }
    time_utils::Timer timer;
    // create 3d clusters from image labels
    std::unordered_map<uint16_t, Cloud> clusters;
    // assign an arbitrary ground label
    uint16_t label = 2;
    // this depth image should contain only ground points
    const cv::Mat& img_ground = cloud.projection_ptr()->depth_image();
    for (int row = 0; row < img_ground.rows; ++row) {
      for (int col = 0; col < img_ground.cols; ++col) {
        const auto& point_container = cloud.projection_ptr()->at(row, col);
        if (point_container.IsEmpty()) {
          // this is ok, just continue, nothing interesting here, no points.
          continue;
        }
        if (img_ground.at<float>(row, col) == 0) {
          // skip non-ground points
          continue;
        }
        for (const auto& point_idx : point_container.points()) {
          const auto& point = cloud.points()[point_idx];
          clusters[label].push_back(point);
        }
      }
    }

    fprintf(stderr, "INFO: prepared ground clusters in: %lu us\n",
            timer.measure());

    this->ShareDataWithAllClients(clusters);
    fprintf(stderr, "INFO: ground clusters shared: %lu us\n", timer.measure());
  }

I am not sure what I miss here so that the processed pointcloud is exactly as the original?

[tuandle]

@niosus any advice, please :)

[niosus]

@tuandle I'm sorry, I am currently really in a complete lack of time to debug your problem. The idea that you have is valid and should theoretically work, but I cannot look for an error in your code right now. I will keep it here and if I find the time I will look into this, but I cannot promise when this might happen. Sorry for that.

[tuandle]

@niosus no problem, thank you for your help anyway :)

[lonlonago]

@niosus the _para._step is 870, does that means for speed up, you do not use all the points into process?

[seboz123]

@tuandle Hi I would like to recreate the same thing: Just printing publishing/ printing the removed ground of the point cloud to evaluate it against other algorithms. Have you found a solution/implementation yet?