Kinfu icp implementation

So i have set if rgbd images which are taken from Microsoft Kinect Sensor V2

KinectFusion: Real-time 3D Reconstruction and InteractionUsing a Moving Depth Camera - https://dl.acm.org/doi/pdf/10.1145/2047196.2047270

In this they have said that the ICP would be done in real time and when using their microsoft app it was nearly real time.

So i used the RGBD odometry methods in open3d i am not getting as speed as the paper suggested.

Is it a different algorthim or am i doing something wrong ?

To be real-time, the reconstruction needs to use CUDA (as they say on the paper)…At the moment, Open3D does not use CUDA… then, it is not possible at the moment…

I ran both algorthims on the same machine and i dont have a nivida GPU

Microsoft app uses DirectX to accelerate, which also uses GPU. We haven’t supported GPU for this task, but it is in progress.

In addition, Open3D’s pipeline, though can be very fast on GPU, is an offline system. It can be very accurate both in terms of reconstruction and trajectory, at the expense that you have to collect all the data before you start to run the system.