Hey, how did you achieve "part" detection and recognition, I mean displaying of the green box? maching learning? would you give me some advice about that? Many thanks.
I actually have three videos in the queue that explain the techniques behind it. There were some complications to get them out. But I have still to wait for a paper to be accepted. It should not take very long anymore since I already work on that for a year. Also, there is a good chance that all of this will become open source at the end of this year. No guarantee, but I work on that too. To your questions. It is a three-step process: Detection -> registration -> tracking. Detection: feature descriptor matching using pairs of Principal Curvatures to match geometry properties, e.g., edges. Registration (and pose estimation): It is just plain ICP. Tracking: Kalman filter. Of course, there are plenty of details that make it work. And the essential steps run on a GPU to maintain real-time. We also work on a CNN solution for detection in point clouds which will replace or substitute the feature detector.
Thanks. Some sort of. We published some aspects but are still short of the entire story due to grant restrictions. But I am working on an open source version + paper (for two years, and I already promised it for two years). But we are close.... 2019 will be the year.
Thanks. The graphic is based on OpenSceneGraph. Detection and tracking are mostly written in C++ (no tracking SDK). We use some support libs such as Eigen, Cuda, and OpenCV for basic functions.
Rafael Radkowski Thanks for ur response. Much appreciated. Okays, it's not totally different from what I thought. I have worked on thingworx. So, i was thinking this is somehow related to AR in IoT .
Hello! Have you published a paper for this work? If so, what is the title of the paper? Thank you!
Hey, how did you achieve "part" detection and recognition, I mean displaying of the green box? maching learning? would you give me some advice about that? Many thanks.
I actually have three videos in the queue that explain the techniques behind it. There were some complications to get them out. But I have still to wait for a paper to be accepted. It should not take very long anymore since I already work on that for a year. Also, there is a good chance that all of this will become open source at the end of this year. No guarantee, but I work on that too.
To your questions. It is a three-step process: Detection -> registration -> tracking.
Detection: feature descriptor matching using pairs of Principal Curvatures to match geometry properties, e.g., edges.
Registration (and pose estimation): It is just plain ICP.
Tracking: Kalman filter.
Of course, there are plenty of details that make it work. And the essential steps run on a GPU to maintain real-time.
We also work on a CNN solution for detection in point clouds which will replace or substitute the feature detector.
great job!
Have you published a paper related to this work? Thanks
good job. i have a question. Where did the Unity engine parts assets come from?
Thanks. Some sort of. We published some aspects but are still short of the entire story due to grant restrictions. But I am working on an open source version + paper (for two years, and I already promised it for two years). But we are close.... 2019 will be the year.
Hi THis looks fantastic. Just would you share which AR platform is used as SDk. Unity3D?
Thanks. The graphic is based on OpenSceneGraph. Detection and tracking are mostly written in C++ (no tracking SDK). We use some support libs such as Eigen, Cuda, and OpenCV for basic functions.
Rafael Radkowski Thanks for ur response. Much appreciated. Okays, it's not totally different from what I thought. I have worked on thingworx. So, i was thinking this is somehow related to AR in IoT .
Is it possible to achieve this functionality through Vuforia?
I think so. With model-based tracking, it should be possible to realize something such as this.
Have u paper? Can ue teach me