Volumetric 3D Mapping in Real-Time on a CPU

Рет қаралды 17,949

10 жыл бұрын

This video shows our volumetric multi-resolution mapping system for RGB-D images that runs on a standard CPU. The fusion of the RGB-D images runs on one CPU core at more than 30 Hz. On a second CPU core, our method generates a textured triangle mesh that it continuously updates at approximately 1 Hz.
Publication:
Volumetric 3D Mapping in Real-Time on a CPU. (F. STeinbrücker, J. Sturm, D. Cremers) To appear at the International Conference on Robotics and Automation (ICRA) 2014.
For more information visit:
vision.in.tum.de

Пікірлер: 16

@jurgodam 10 жыл бұрын

Hi guys, I'm glad you like the video! Just to clarify this, we assume here that we already have the camera poses (in practice, we get them from Christian Kerl's dvo system). This paper is about how to generate a volumetric model (given the camera poses) and how to generate a mesh out of it, in real-time, on a CPU. We plan to release the source code of this, stay tuned!

@JavierCivera 10 жыл бұрын

Great work Jürgen, congrats! :D

@constructiverealities 4 жыл бұрын

this method looks very interesting and newly relevant with faster multi core ARM processors in 2019. was this code ever published? very interested in learning more or trying to run it on current day hardware.

@constructiverealities 4 жыл бұрын

found it!

@mattizzle81 5 жыл бұрын

Still state of the art in 2019. I haven't found anything better, nothing else can do this on CPU. At least nothing open to the public.

@RobertGreen0103 10 жыл бұрын

Love the work, and once again the source code is available! You guys are just pushing the sate-of-the-art forward!

@aitjellal 10 жыл бұрын

amazing !!

@riteshpatel1460 8 жыл бұрын

can we do this with monocular rgb camera ??

@jonasls 10 жыл бұрын

How do u do this?

@MrHannatas 9 жыл бұрын

can this code run on a ARM Powered device ? (i am looking for a kinect volumetric aquisition software that can run on a Kinect) Speed is not that important at the moment and the RPi will run at 1000Mhz

@mattizzle81 5 жыл бұрын

I know this is a very late reply, but I was able to get it to compile on arm using sse2neon. It does not compile for arm unmodified because it uses intel specific instructions. I got it running on a jetson nano and on a samsung note 9. However my problem is it ends up eating tons of ram and ends up slowing down rapidly due to swapping, using more than 8GB of ram on both devices. That makes it unusable.

@constructiverealities 4 жыл бұрын

@@mattizzle81 very interesting. wondering if newer SOMs with PCIe/nvme storage can improve the swap speeds.