Using CNNs for Handbag Brand and Color Detection | Hacker News Reader

Using CNNs for Handbag Brand and Color Detection

technology.condenast.com

97 pointsgautamsarora21 comments

Although they describe the process of how they went about collecting the data set, it would have been an extra special touch for them to release a formatted/clean version of it!

Wow. It's really crazy how it can pick the handbag out even with complex similar backgrounds.

I don't quite understand the object localization piece of the article. Are there other resources on this topic that explain how do do the same thing (object localization). Or tutorials/examples on this topic?

There are quite a few papers that discuss the application R-CNN (Region-based Convolutional Neural Network) for object detection and classification [1-3]. On websites such as Medium, you can find lots of blog posts on object detection using popular frameworks such as Tensorflow (e.g. https://towardsdatascience.com/how-to-train-your-own-object-...).

[1] Faster R-CNN: https://arxiv.org/abs/1506.01497

[2] Single Shot Multibox: https://arxiv.org/abs/1512.02325

[3] YOLO9000: https://arxiv.org/abs/1612.08242

It's based on this paper https://arxiv.org/pdf/1512.04150.pdf The approach is different than the Faster R-CNN etc mentioned below as it is not fully supervised, i.e. the position of the objected is not in the training data. I can write a simple tutorial and link in the article and here later today.

I should mention that the accuracy is lower than the methods mentioned below.

> I can write a simple tutorial and link in the article and here later today

That would be fantastic. Thanks!

I am curious if you can use this "process" to spot a fake. Fake handbags are a BIG problem in the luxury accessories market. R-CNN. Object detection and classification.

A big piece missing here is data augmentation. If they had more of it maybe they wouldn’t need to do the separate final layer training.

I wonder if it will ever get good enough to detect counterfeit bags? My understanding is that they are a prolific problem.

Maybe a certain percentage, but much of the work in detecting many counterfeit bags involves actually inspecting a lot of the work on the interior of the bag or looking for specific construction characteristics that knock-off groups don't take the time to replicate for a market that is unknowing.

source: a mom that was nuts about calling out fakes in a snarky manner.

My thinking was that one might be able to do it with their cell phone, with close up images. Then it could maybe tell things like stitching, material, precise color match, and maybe then be able to say if it is counterfeit.

I'm just not sure ML is yet that good. My thoughts were more about multiple pictures or a small video. I'd suspect that it could train itself further as the DB expands.

I'm just not sure how good it actually can be. It probably only has to be better than a layperson, though I could see more precise stuff used to scan by customs. Any percentage higher than the normal buyer of such would be better, assuming few false positives.

You could certainly do that, but you would need to create the training dataset. It would require many thousands of photos of real and counterfeit bags up close.

Image recognition has been mostly solved by ML techniques, so if you just assemble the data and train the network you can distinguish any set of things that are significantly different visually.

You shouldn't be getting downvoted for this.

Yes, this is an active area of research. There was a paper at KDD17 working on this: https://dl.acm.org/citation.cfm?id=3098186&dl=ACM&coll=DL&CF...

NVidia had a blog post about this: https://blogs.nvidia.com/blog/2017/08/03/detecting-counterfe...

There's a French company with a product: http://cypheme.com/

Neat, thanks! I could see a few valuable use cases, from customs to personal.

I have no idea why condenast got into tech, I believe it is simply getting trendy with their crowd

Presumably they're using this to sell handbags. Like to put "buy now" buttons under images on their properties or something.

Hi - Team Lead from Condé Nast here. You're right in that WIRED and New Yorker readers are interested in ML, but that's not solely why we've invested. This work is driven by our platform research team (FORE) which produces predictive models for a [wide] variety of uses, primarily to deliver a better user experience for our readers.

For example, these models are directly integrated into our search, recommendation, SEO optimization engines. Those engines then power both our in-house CMS as well as the modules you see on the brand websites.

I wonder how long it took to train.

I took about a day on one aws p2.xlarge (one Tesla K80) starting from a pre-trained inception model (trained and opened sourced by Googled on ImageNet data).

Killer work Johan and Paul!

Reply on news.ycombinator.com