Analytics Claims To Automatically Detect Car Make / Model

HP Autonomy (the company whose acquisition was a disaster with claims of fraud), is back in the surveillance industry marketing analytics.

Their new pitch is that they can automatically identify the make and models of vehicles (i.e., that car was a Ford, this one is a BMW, the car at 10:15 was a civic, etc.).

Here's a screencap from their demo video:

And here's the full demo video itself:

What do you think? Would you use it if it worked? Assume it's going to be expensive (obviously).

The two big things that jump out at me immediately is how incredibly tight the FoV is and how difficult I'd expect this to be in various lighting and weather conditions as there are lots of makes and models of cars that can look similar.

I have seen software recognizing emblem at the front and back of the vehicle to match the manufacturer, so that is definitely do-able with high speed cameras catching a good still image, but saying it can tell the model of the make is far fetch with such FoV, even with mult--cameras environment, i am still skeptical about the accuarcy as it will really requires a huge CPU processing capability for such algorithm.

Even if it does work in non-ideal lighting conditions (doubtful), how are they hoping to recognize cars with any modifications including removed emblems (intentionally or not)??

Fancy sounding marketing tool.

Seems unlikely to me. This doesn't follow with the logical progression of technology. There are a small number of analytics products today that can accurately distinquish that an object is vehicle with decent reliability. None of these are perfect, but some are close enough that the false positives and false negatives are low enough to be deemed acceptable. Still, you're talking about specialized products at price points that ultimately limit their deployment options from being included by default on all products/channels.

This demo attempts to say that before we've reached a point of mainstream reliable analytics, that we've already moved on to making nuanced decisions about the objects detected. And furthermore, from a company that hadn't been a forerunner of video analytics to begin with.

"And furthermore, from a company that hadn't been a forerunner of video analytics to begin with."

So I agree with you. However, their counter is going to be that they are the grandmaster / pioneer of analytics in general. I remember that was their pitch to me 9 years ago (at ASIS Orlando 2005) when they made the same claims. It truly is deja vu.

But they won an award. They couldn't have won an award if it didn't work.

"But they won an award. They couldn't have won an award if it didn't work."

Cut it out, Carmona! The judges spent 15 minutes evaluating this, which is more than enough time to verify that it works....

Forgot to link.

Awards, Shamwards...

If they can just read the plates, can't they cross reference vehicle registration anyway?

If you're looking for a blue 1997 Honda hatchback, can't you just flag the plate numbers all registered to that type of vehicle to be 'hot' or something?

I'm sure stolen plates/removed plates are still a problem, but that's why you have more than fancy analytics on your side.

Most Motor Vehicle databases only have the manufacture of vehicle and possible truck or sedan clasification, not the model and color. So the ability to cross reference ALPR with the usual state data is rather limited.