Nvidia's demo of real-time object recognition using deep learning [video] | Hacker News Reader

Nvidia's demo of real-time object recognition using deep learning [video]

youtube.com

75 pointsvkhuc23 comments

The CEO butting in all the time was really annoying, had a business partner that did this in meetings all the time - its fucking annoying and rude. On the surface I have no idea if this is ground breaking or not, my first thought was ahh nVidia using Linux!

Yeah, and his colour commentary is subtly wrong, too.

Maybe because there was no rehearsal.

Unfortunately these are the ones that make it to the top and are named CEO.

Looks great! Can anyone with more knowledge about deep learning say whether this is an exceptional achievement in the field ?

As far as state-of-the-art classification goes, this is not terribly impressive. See the recent results in something like:

http://cs.stanford.edu/people/karpathy/deepimagesent/

I think the impressive thing here is that the GPU is presumably doing GIANT matrix multiplications in real time. A prediction from a neural net is just a series of matrix multiplications, and matrix multiplications are about n^2.8 in complexity, so you can see how matrix multiplications with thousands of rows/columns (often what these sorts of deep image classifiers involve) are hugely computationally expensive.

So it's definitely important for real time machine learning systems to have access to this kind of linear algebra power, but the actual machine learning techniques demonstrated are not super impressive. The hardware is. Which makes sense since this is an Nvidia demo.

From what i can understand what's even more impressive is that it was running on a beefed up version of their latest mobile SOC and not on some 5000$ compute GPU card. Which means that this application can be both very affordable and very practical since people won't put a 300W GPU in their car.

Definitely agreed. When well-maintained and easy to use machine learning libraries meet very powerful, highly embeddable GPUs (or other dedicated linear algebra sort of hardware), I think we'll see a big revolution in the whole "smart object" field.

Right now you have iPhones and whatnot doing touch ID with fingerprints, but imagine if your phone could recognize you just by quickly analyzing the gyro data as you raise your phone and comparing it against the other thousand times you've pulled your phone out of your pocket, combined with the slight pressure readings near the touchscreen's edge because it's learned where you're fingers fall on the case.

^ contrived example I just thought of, but you get the idea.

I suppose that training the network requires the cluster stuffed with GPU's?

Intel 'realsense' drone demo was, and I'm not into the smart/IoT trend, somehow impressive. A flying electromechanical bug on stage at a mainstream show, to me that was a small but real inflection point.

https://www.youtube.com/watch?v=Gn83Psbv61I

ps: I'm not sure it was fully real-time though, the door avoidance restart seemed a little too nicely cued.

Yeah those numbers aren't particularly accurate. GTX 980, which costs about $600, has 2048 cuda cores and 5 tflops peak, with 150 watts consumption.

Yes because some where in my sentence i was referring to a 980 or a 780ti some where? Maybe it was a Titan Z, no maybe it's still the GTX 690 Ti which is still the fastest single card they made, or maybe that statement was referring to their latest COMPUTE card the Tesla K80 which costs 5000$, and contains 2 new Kepler cores (GM210) and requires about 300 watts of juice to run....

Current state of the art is a bit better than this. See the bottom section of [1] for some of the latest publications.

However, building a real world working system has challenges that are different to the academic challenge of trying to classify the most classes possible in static images.

[1] http://blogs.technet.com/b/machinelearning/archive/2014/11/1...

A few commenters already tell that this isn't really groundbreaking work, but how about for real time? And commodity hardware (this one is a few watt mobile chip)?

I'm not an expert but, from what I do know, it seems like the take-away here is that it's running on tech which is within reach of most consumers. Sure, academics are accomplishing more impressive feats in the lab. However, it looks like nVidia has brought these algorithms onto hardware which is probably not much different from what they're already selling to gamers at a reasonable price. That could end up being a really big deal and could boost applications of computer vision in consumer tech significantly.

Here's how to do the street sign part of this yourself: https://gist.github.com/iandees/f773749c47d088705199

Cool demo but I still wonder if fundamentally this is just a brute-force approach. Wouldn't it be better to do some traditional preprocessing (e.g. recognizing rectangles, circles, etc.) and feeding higher-level descriptors into the classifier?

If the net learns based on pixels you still have to somehow solve rotation and scale invariance. Or is there something new in deep-learning vs. old-school neural nets that fixes the issues that bedeviled neural nets the first time they were popular?

I think they used the methods described in http://www.cs.berkeley.edu/~rbg/papers/r-cnn-cvpr.pdf

Thanks, interesting paper.

video shows demo happening in ubuntu --- at least the video playback

@10:08

on the right merc sls classified as SUV

on the left one SUV classified as two VANs

Their algorithm works at about 1Hz rate when doing signs. This is ~state of the art from 20 years ago, but running on small mobile SoC at a slow rate.

Please show a paper where fine-grained vehicle classification in unconstrained images is anywhere near this performance from 20 years ago. You will not be able to, because it wasn't.

state of the art classification accuracy/range, not speed.

Reply on news.ycombinator.com