Deep Learning

Can we use Deep Learning to improve object recognition in AR applications for maintenance?

CH Asked by Christopher Stevens · 10-01-2023
0 upvotes 16,863 views 0 comments
The question

We are building an AR app to help technicians identify complex wiring in server rooms. Standard image tracking is failing due to low light and similar-looking cables. Would integrating a YOLOv8 or similar Deep Learning model improve the accuracy of our AR overlays in these environments? 

3 answers

0
LI
Answered on 15-03-2023

Absolutely. YOLOv8 is excellent for this because it’s optimized for real-time edge devices. The challenge with server rooms isn't just the light; it's the "Visual Noise." You’ll need to train your model on a very specific dataset of your actual server racks in various lighting conditions. To make this work with AR, you’ll need a bridge—usually something like ONNX Runtime—to run the model on the mobile device while the AR framework handles the spatial anchoring. It’s a heavy lift for the processor, so you might need to run the inference at 5-10fps while the AR tracking runs at 60fps to save battery.

0
J
Answered on 22-04-2023

Linda, that sounds like a battery killer. Instead of full-frame Deep Learning, could Christopher use "Semantic Segmentation" just to isolate the wires and then use traditional edge detection for the actual tracking? Wouldn't that be more efficient for a handheld device that needs to last a 6-hour shift?

PA 30-04-2023

James, you're onto something, but "Semantic Segmentation" is still quite compute-heavy. A better middle ground might be using the DL model to "re-localize" the position every few seconds while using the IMU (gyroscope/accelerometer) and SLAM for the high-speed tracking in between. This "Hybrid Tracking" approach gives you the accuracy of Deep Learning without draining the battery in twenty minutes.

0
KA
Answered on 10-05-2023

Don't forget the training data! For server rooms, you should augment your dataset with "synthetic data" showing different cable colors and tangles to make the model truly robust. 

Synthetic data was a game changer for us. We generated 5,000 variations of "messy cables" in a 3D engine to train our model, and the real-world accuracy jumped from 60% to 92%.

Share your thoughts

Your email address will not be published. Required fields are marked (*)

Professional Counselling Session

Still have questions?
Schedule a free counselling session

Our experts are ready to help you with any questions about courses, admissions, or career paths. Get personalized guidance from industry professionals.

Request a Call Back

Search Online

We Accept

We Accept

Follow Us

"PMI®", "PMBOK®", "PMP®", "CAPM®" and "PMI-ACP®" are registered marks of the Project Management Institute, Inc. | "CSM", "CST" are Registered Trade Marks of The Scrum Alliance, USA. | COBIT® is a trademark of ISACA® registered in the United States and other countries.

Book Free Session