When Jonathan Tapson presented the Roadmap, he foreshadowed a couple of inventions in the pipeline which would improve efficiency (to do with moving data from memory).
This new BRN web page discusses Akida's CenterNet model, a "memory-friendly object detection model":
The Akida CenterNet model delivers production-ready object detection capabilities with unprecedented efficiency for edge deployment. Unlike full-size CenterNet, Akida’s CenterNet model is compressed and quantized, enabling real-time detection at low power with a tiny footprint.
1. Capture
- Input images through lightweight, optimized CNN backbone
- Extract essential visual features while maintaining computational efficiency
2. Process
- Refine features through compact neck layers for optimal detection
- Generate heatmaps, size predictions, and position offsets simultaneously
3. Detect
- Identify object center points from heatmap peaks
- Combine predictions into precise bounding boxes around detected objects
Ultra-Compact Design
Lightweight and memory-friendly object detection model for the smallest devices without compromising functionality or performance.
PS: I'm guessing the patents have been filed.
Last edited:
. I have 7 on my ignore list; therefore, I cannot see his posts. Without context, your post did not seem nice.
In short: I realistically see MB.OS as a hybrid – Nvidia as the main computing platform, plus additional chips like Akida for edge learning. Mercedes won’t market it as “Akida inside” but rather under broader labels like “intelligent personalization” or “hyper-personalized experience”.
