How to copy sensor_msgs::PointCloud2 to tensorflow::Tensor fast?

asked 2020-02-13 06:49:53 -0500

danilobr94 gravatar image

Hi,

i am trying to copy a sensor_msgs::PointCloud2 into a tensorflow::Tensor.

Currently, it is done like this:

void pclCallback(const sensor_msgs::PointCloud2 &pcl2_msg){
pcl::PointCloud<pcl::PointXYZI> cloud;
pcl::fromROSMsg(pcl2_msg, cloud);

const int number_of_points = cloud.points.size();
tensorflow::Tensor points_tensor{tensorflow::DataType::DT_FLOAT, tensorflow::TensorShape({1,number_of_points,4})};

auto pointsMapped = points_tensor.tensor<float, 3>();
for(int i=0; i<number_of_points; i++){
    pointsMapped(0,i,0) = cloud.points[i].x;
    pointsMapped(0,i,1) = cloud.points[i].y;
    pointsMapped(0,i,2) = cloud.points[i].z;
    pointsMapped(0,i,3) = cloud.points[i].intensity /255.0;
}

The performance is quiet bad (both for the for loop and the pcl::fromROSMsg(pcl2_input, cloud).

Is there a recommended way to do this?

edit retag flag offensive close merge delete

Comments

You could reduce cost using pointers in both PCL ans Sensormsgs point cloud types

Fetullah Atas gravatar image Fetullah Atas  ( 2020-02-13 07:43:30 -0500 )edit