43 comments

[ 3.2 ms ] story [ 110 ms ] thread
For the threshold value, how is 0.25 determined to be the most optimal? is this value just specific to the datasets YOLO was trained on or is it something that is universal.

Also, this is really awesome!

Generally you learn a threshold value like that under cross validation.
(a) very cool project

(b) ..interesting CV

If you look closely at the video, the quality of detection is not really great for anything besides `person`. I wonder if weighing objects by the rate they are being found over N frames, and conforming to a somewhat linear trajectory could improve results; feeding matches from previous frames could also help find the same object in subsequent ones, and maybe even work out temporary occlusion (thinking out loud here - probably way more complex than one would expect).
I'm not 100% sure about this but I think part of the reason why its called you only look once is that it is single frame object detection. I do know that there are other types of networks that can previous frames into account, but the models themselves are much larger (in terms of VRAM requirements for loading into memory) and more computationally intensive. This network is special because it can run with very little power. From what I remember it can run on a 10W board at 6fps which compared to networks only a few years ago is 10x decrease
It's clearly single-frame; you can watch recognition succeed and fail from frame to frame. The "person" recognizer misses some clear faces, so it's not heavily face-oriented. Since it's single-frame, it's not recognizing articulated motion.

Recognition seems to be limited to "person", "motorbike", "tie", "cell phone" (a gun, actually), "umbrella", "truck" (misidentified part of a train) "bench" (a railing) and "horse" (motorbike with duffel seen from rear). "Person", "umbrella", "tie", and "motorbike" seem to work; the others are kind of random.

The trouble with running recognizers on Hollywood movies is that they have many conventions of what appears on screen and how big it is on screen. Are they training on such data?

Good data sets would be side views from a moving vehicle, like Google StreetView data or just a GoPro pointed sideways while driving around.

It still seems to pick up a lot more objects than some of the video we've seen from, eg. Tesla Autopilot.
Hypothesis: All demos are improved with the addition of dragonforce.
Discovery: Blaring music of any kind, including my preferred genres, is unpleasant when trying to watch a demo.
Doesn't hold up in independent tests.
most of the top submissions to the last ImageNet comp. are tweaks on YOLO and Faster RCNN.

But why is this trending now?

could be because more people are getting into DL and image recognition? I am one of them so what might be old to some people is new info for beginners
This is good background but modern architecture are significantly better:

Learning Transferable Architectures for Scalable Image Recognition

https://arxiv.org/abs/1707.07012

This is not "modern" though, the paper is quite old. He has improved it significantly since then: https://arxiv.org/abs/1612.08242. Best paper honorable mention at CVPR2017.
I missed that the site was updated for YOLO9000, it is very impressive. The Google paper is recent though (July), maybe you're thinking of https://arxiv.org/abs/1611.10012? Anyway, what I meant to point out was that there has been a lot of work since on improving the efficiency of the architectures used for feature extraction (Xception, MobileNet, ShuffleNet, NASNet).
Sure, there's new work all the time. But I think you might be missing the forest for the trees here. The main efficiency gain here is not in how you extract and transform features (although those improvements would also work in YOLO, as far as I can tell). The novel part is in structuring the problem as a single forward pass regression. Further improvement in this year's paper is in a better loss function. Other possible efficiency improvements are orthogonal and complementary to that.
I agree that the single shot approach in YOLO was pretty clearly a big step forward. In the two years since though there much of the efficiency work has been in the underlying feature detector architecture, which as you point out should integrate well with the YOLO9000 training improvements. It's very cool to see this kind of capability get within range of phone-sized devices.
This is cool but outdated, there're other better papers now.
What better papers would you recommend reading?
Arxiv is thou friend :)

https://www.google.com/search?q=fast+object+detection+deep+l...

and

https://www.reddit.com/r/MachineLearning/search?q=object+det...

There's been an absolute deluge of papers, I can't say I've kept up with them all. There was one interesting one in particular was able to learn object detection in an unsupervised manner in a novel improvement upon Bottou's "is object detection for free" paper.

> There's been an absolute deluge of papers, I can't say I've kept up with them all.

That's why recommendations were asked for.

To save someone googling, the referenced "Is object detection for free" paper is at http://leon.bottou.org/publications/pdf/cvpr-2015.pdf

I also just came across "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks", which looks interesting and is at https://arxiv.org/pdf/1506.01497.pdf

(I'll be reading these after work tonight.)

These are both older and not as good as YOLO9000.
Nice.

How good is face recognition at the moment?

Or is this a completely different topic :D

YOLO9000: Better, Faster, Stronger won an honourable mention award at CVPR in June. There maybe better papers, but they are very recent: the last six months taking lead times into account.
Nice! Btw this guy's CV is awesome :) : https://pjreddie.com/resume/
Whoa that was a download link

(On Android)

Chrome on Android auto downloads PDFs.
Odd I'm in firefox. I think Chrome for me asked which app to open with but that probably means it was already downloaded.

I don't know I've heard of malware with PDFs, I know probably a stretch on Android and this person/HN probably not really a concern. Still don't like auto-download unless expecting one. But I get it was the lack of a browser PDF displayer. Maybe he isn't running PDF.js or whatever that is you run to display pdfs on your site.

(Sorry for so much text, TL;DR I'm afraid of what I don't know) I think I'm secure, but am I really secure?

I wouldn't hire someone who formats their CV like that. Employment is a serious proposition, and ... Ah just kidding. Nice CV.
It's the kind of thing that you can only get away with if you're very, very good. At that point it becomes a kind of power move, like showing up to a high end interview in jeans and a t-shirt.
Or if you pretend that you are very good. If you show up to an interview in jeans and a t-shirt, people will think you are very good.
True, it can also be a form of peacocking. But in an industry where results are easily tested it's much harder to get away with that.
In this case, the guy has been spearheading a very high performance, innovative methodology for real-time object detection. Companies interested in that task should be falling over themselves to hire him.
Yea, that resume takes a certain amount of chutzpah, doesn't it? Kind of a clever idea for a cover letter, though, with a "standard" resume attached.
Impressive. Does anyone actually know of real apps that make use of object detection? Apart from Google apps.
Not hotdog?
haha. Had no clue the guys at Silicon Valley built an actual app for that. Thanks for sharing.