RE: MOU semiconductor10 Nov 2020 18:45
Finally able to post this on LSE
In the Ali Osman article, NXP were trying too hard, all that talk of TensorFlow and Caffe were like a 4 year old saying "look Mummy I can use grown up tools".
NXP are a European company, but all the DMS pictures were Chinese – that tells you the source and the intended destination. A young company from the boom times of AI and ML – we can do anything with ML, just give it the annotated data and it will learn better than a human. Well that is true, up to a point. Many students could quickly knock up a DMS system that works in the lab 50% of the time add some more money and good data and you can get funding when you improve the performance. Perhaps you get it to 90% but that is when it matters, the next 5-10% will take 90% of the money and effort! Seeing Machines did it the other way around, the base is procedural, with AI learning on top, and it was fed 5 billion km of real world annotated data, so they have the edge cases covered, there availability and accuracy is far beyond what these start-ups can do and they have the IP to prove it.
The source is Momenta AI and they were founded in Beijing in 2016
https://www.momenta.cn/en/apply.html?id=mpilotbox
So NXP will be good enough for China today, it may even end up in some European cars at the low end, after all it ticks a DMS box and it comes with all of the other features, but the market will soon out them. Customers, OEMs and Regulators will force them to upgrade. But for now it shows that there is a rich and varied competition in the various spaces, so we are not a functional monopoly.