https://arxiv.org/abs/1603.05279 is the research paper on which the product is based. They have optimized it to the core. Any comments on how this compares with SSD/Yolo?
Anuragchil this paper is from 2016, RetinaNet is from Aug 2017, and they don’t even mention xnor publication so I would guess they don’t have anything that is close to state of the art.
Are you sure the company’s tech is based on that paper? There doesn’t seem to be much info on their web site…
The xnor paper shares 2 authors with the YOLO papers. It’s a really interesting paper that suggests a path towards very low-power chips for computer vision tasks, IMO.
The VP of engineering mentions that they “optimized some efficient algorithms like xnor net to the hilt” in his linkedin profile - https://www.linkedin.com/in/dmitrybelenko/. Any idea if there is an implementation available in pytorch? I couldn’t find any.
I didn’t see a lot of discussion of that paper after it came out - it seemed to get a little forgotten. I’m not aware of a pytorch implementation. I’m glad a startup is trying to take it further, although it’s a shame it’s not open source.
@jeremy they just secured a $12million series A. They’re definitely not forgotten, but their operation is definitely hush hush. Yesterday(or a couple of days ago) they hinted at their business model - Selling models that are optimized to be trained/embedded in devices to perform CV tasks on edge devices(drones, security cams, etc). They’re still very early stage, so no indications yet on mass availability, but they’ve been garnering good press at least here in the Seattle area, and depending on accuracy levels of their binarized models, they’re an interesting startup to follow! Here’s their updated website! https://www.xnor.ai/
Their domain is down.
Apple acquires Xnor.ai, edge AI spin-out from Paul Allen’s AI2, for price in $200M range
Apple buys edge-based AI startup Xnor.ai for a reported $200M