tagging.tech

Audio, Image and video keywording. By people and machines.


Leave a comment

Tagging.tech interview with Martin Wilson

Tagging.tech presents an audio interview with Martin Wilson about image recognition.

 

Listen and subscribe to Tagging.tech on Apple PodcastsAudioBoom, CastBox, Google Play, RadioPublic or TuneIn.

 

Transcript:

 

Henrik de Gyor:  This is Tagging.tech. I’m Henrik de Gyor. Today I’m speaking with Martin Wilson. Martin, how are you?

Martin Wilson:  I’m very well, thank you. How are you?

Henrik:  Good. Martin, who are you and what do you do?

Martin:  I am a director at Asset Bank. Being a director, I’ve done an awful lot of different things over the years. I have done some development on our product, Asset Bank. I’ve done sales and I’ve done consultancy while rolling out the product.

Just to explain a little bit about what Asset Bank is as a product, it is a digital asset management solution. Digital asset management is often shortened to DAM. A DAM solution helps clients and the users to organize the digital assets that almost every organization owns and makes use of nowadays.

By digital asset, we mean primary files. Things like images, videos, documents and all of those. A digital asset has an awful lot of value to an organization and it’s very important that they can find them easily, that they don’t waste money recreating digital assets that they already have, and that the assets themselves are used properly in a way that’s consistent with the brand of the organization.

Henrik:  Martin, what are the biggest challenges and successes you’ve seen with image and video recognition?

Martin:  Let me first start by saying how I think that image recognition has a potential to have a really big impact on my industry, digital asset management. Digital asset management is all about being able to find images and then use them properly. That’s the purpose of the DAM system. There’s an old adage which people use and it says that a DAM system is only as good as the metadata that is associated with the assets. The reason for that is, a million images, if you have a million images in any system it’s almost impossible to find the image you want without some sort of a search and or a browse function. Those searches and browse functions at the moment rely on what we call metadata that it is associated with the assets. That metadata is things like title or caption of an image, description, perhaps some keywords that been put in, maybe some information about how that can be used, the image can be used.

The result of this is that people, humans, spend an awful lot of time entering the metadata that is associated with digital assets. Usually, within an organization, the processes, the workflows that are associated with using a DAM application involve uploading one or more or many digital assets, typically images or videos, and then manually entering the data by, for example, looking at the image, seeing what it’s about, what the subject is, who’s in it maybe if it’s of people and then just actually typing in that data.

As you can imagine, that takes a lot of time. It’s also considered quite boring by most people. For that reason, it’s often skipped or not done really well. If it’s not done really well, the data associated with the assets is incomplete and therefore it’s very hard for it to turn up in the right searches.

The idea that it could be automated, this process, and have a computer work out what’s in the image and tag the digital assets appropriately is enormous. It’s almost like the Holy Grail of the upload process for DAM systems.

There was an awful lot of excitement when, for example Google Cloud Vision came out with their service. It’s what called an API which enables other applications to make use of the image recognition functionality. There’s a lot of other services as well that have come out in the last couple of years like Clarify, is another one.

When they came out, lots of DAM vendors got very excited and rushed to add the functionality into their own applications. We did the same. About a year ago we started a project with the objective of developing a component that could be used with Asset Bank in order to add auto-tagging capabilities to asset bank.

Let me just describe some of the challenges then that we found in doing that and when we rolled out some of our clients, the challenges they found. One of the challenges, I suppose which is always like a umbrella challenge over all of it, is people’s expectations.

Humans are very good at looking at images and working out what’s in it. They’ve also got a lot of domain knowledge. Usually, they understand, for example, their products. They can look at a product shot and say, “Yeah, that’s product F-567”, or whatever the code is. It’s actually very hard for computers to do that well. That problem hasn’t been solved that well yet.

What we found is, when compared with how humans tag images, the results coming from the auto-tagging software or APIs was not, to be frank, not of good enough quality for most cases. That’s the second specific challenge then, really. The quality of the raw results coming back from the software. The image, the visual recognition software was not quite good enough for use in most organizations, especially in a commercial sense.

That’s not say that it’s not useful. I’ll come on to that in a bit. What we found, on to the successes, what we found was that certain clients who had more generic or general images, the results were much better. We’ve got some clients who are tourists boards. They’ve got images of landscapes and scenery. Most of the image recognition software is quite good at finding the subjects and suggesting keywords for those types of images.

One of the reasons for that is that most of them have been trained on image data sets, that are images that are found on the internet for example. Of course they’re going to be generic. The other end of the spectrum, where we found it didn’t work that well was for clients that have got quite bespoke business domains or subject domains, images of their own product range. Very hard for these fairly generic image recognition software APIs to be able to come up with the right keywords for those sorts of image.

That’s possibly where there are still gaps. That might be something we’ll talk about in a minute about the future, which is the inability for a lot of this tagging software to learn from bespoke data sets.

Henrik:  Martin, as of December 2016, how do you see image and video recognition changing?

Martin:  I think it’s fair to say that it’s in it’s infancy at the moment. It’s only since it’s become available through the online cloud services or web services that people have found it very easy to start using this technology in their own applications. It’s only been the last couple of years, that really has kind of taken off as something that can be openly or easily used.

Now I think the vendors of this sort of software are learning very quickly from real use cases. I think it’s quite an exciting for where the commercial or non-commercial application of this software can go. I think if we first focus a little bit more on the current problems, that gives some insight into where the software might go, what direction it might go in.

I was just talking then about one of the problems being that is very generic at the moment, the tags that you get back from the online services are going to be fairly generic. That’s obviously the case if you understand how they work and how they learn. I think very quickly we’re going to see these services, and I know some are already, offering the ability for you to train them with your own data sets. That then opens up the application a lot more widely.

One of the things that image recognition and artificial intelligence, in general, is the context in which they’re operating. It’s much easier for image recognition software to work well if it is working within quite a narrow context. As an example, if you’re talking about, or if you want to try and get the software to recognize your product range, then if it’s trained on images that are of product range, and therefore the context is only products within your product range, then it’s a lot easier for it to recognize the right products, rather than having to think of every product that it’s ever seen an image of in the entire world.

Just to reiterate that I think the ability to train the software in bespoke data sets and for it to concentrate on in effect, domain-specific subjects, I think that’s a must and that will start to happen.

I think we will see quite a few hybrid solutions. What we found when we were doing our investigation into the software and what we ended up doing within Asset Bank or within the components which we call QuickTagger that works with Asset Banks is, coming up with hybrid human and computer interaction model where the tags that were being suggested by the visual recognition software were not just accepted as that’s job done. They were used to then group the images so a user could very quickly change the tags that weren’t right.

They could, for example, accept some of the tags, because they’re the right tags and the human agreed with the computer in effect, but then they could quite easily change the tags that were wrong. The key thing here is that the grouping was still being done pretty successfully. Although the tags that the image recognition software was suggesting might not be right, it was recognizing that certain images were of the same subject. That therefore meant that a human could go in and say, “Okay, I’ve got 50 images here that are all of a particular, I don’t know, model of car. They’ve all been grouped together, so that makes it really easy for me as a human to now type in the right name of the car or the model of the car.”

I think that idea, that where we are right now with this technology is that can help facilitate, speed up the human interfaces. That’s a quite a powerful idea I think, but where … I think that will continue, so we’ll see an evolution of that. I think we’re quite a long way off just being able to say, “Okay, you get on with it, computer. Tag these up.” I think we’re going to see improvements and sort of evolution of the idea of humans and computers working together in this auto-tagging sphere.

Henrik:  Martin, what advice would you like to share with people looking at image and video recognition?

Martin:  The first thing I would say is about expectations management. If you are used to having tags generated by humans who know what they’re doing, they understand the domain, the subject domain of the images that they’re tagging, you are likely to be fairly disappointed I would say in the results for most cases.

That’s one thing. See beyond the raw results you’re getting back from the tagging software. Look to how you might use the tags though to your advantage. For example, in hybrid solutions.

Consider what subject matter you’ve got, what your images are actually of and tailor your expectations accordingly. If you’ve got a lot of images that are of fairly generic subjects, you might find a lot of value from the auto tags. If you’ve got quite specific subjects, be prepared to potentially be a bit disappointed and or to have to put in quite a lot of work to either start training some of the software that you’re using or looking at how you can sort of augment the results with human interactions.

Sorry, another bit of advice is shop around. Have a look at the different services that are available. They’re fairly different. We built our QuickTagger in such a way that we can plug in the different services that are available, so we could just simply change it to work with Google Cloud Vision or with Clarify and there’s ten other potential candidates that I could list off the top of my head and probably more out there. They give different results. Some of them are better for different applications as well and different subjects. Usually, very simple to get a free trial and try out the software that’s there. That would be my last bit of advice. Shop around with the auto-tagging technologies that are available.

Henrik:  Martin, where can we find out more information?

Martin:  More information on our product Asset Bank is available on our website, which is www.assetbank.co.uk. If you’re interested in particular in how we, in the experiments that we’ve done and the components that we’ve got for Asset Bank, QuickTagger, then just fill in our contact form and express that interest. I would personally be very happy to talk to people about what we found.

There’s some information about QuickTagger that we’ve developed on our website as well. If you’re interested in finding about the different technologies that are available out there for you to use within your own application, there’s a lot. Personally, I would recommend now the cloud-based ones, because it’s much easier to get up and running with those. There’s quite a lot of information, meaning if you just typed in ‘image recognition software’ or ‘image recognition APIs’, you’ll see there’s quite a few good articles that people have put together on Quora and so on that have done the research for you. Use that as a starting point because as I say, things change all the time. New APIs come out. Do your research, but there is a lot of information available on the internet about this.

Henrik:  Thanks, Martin.

Martin:  You’re welcome.

Henrik:  For more on this, visit Tagging.tech.

Thanks again.


 

For a book about this, visit keywordingnow.com

 


Leave a comment

Tagging.tech interview with Jonas Dahl

Tagging.tech presents an audio interview with Jonas Dahl about image recognition

Listen and subscribe to Tagging.tech on Apple PodcastsAudioBoom, CastBox, Google Play, RadioPublic or TuneIn.

Transcript:

Henrik de Gyor:  This is Tagging.tech. I’m Henrik de Gyor. Today I’m speaking with Jonas Dahl. Jonas, how are you?
Jonas Dahl:  Good. How are you?Henrik:  Good. Jonas, who are you and what do you do?Jonas:  Yeah, so I’m a product manager with Adobe Experience Manager. And I primarily look after our machine learning and big data features across all AEM products, so basically working with deep learning, graph-based methods, NLP, etc.

Henrik:  Jonas, what are the biggest challenges and successes you’ve seen with image recognition?

Jonas:  Yes. Well, deep learning is basically what happened, what defines before and after. So, basically in 2012, there’s a confluence of the data piece that is primarily enabled by the Internet, large amounts of well-labeled images that could drive these huge deep learning networks. There’s the deep learning technology and, obviously, the availability of raw computing power. So, that’s basically what happened. And with that we saw accuracy increase tremendously, and now it’s basically rivaling human performance, right? So we see both accuracy and also kind of the breadth of labeling you can do in classification you can do has just increased and improved tremendously in the last few years.

In terms of challenges, what I see is, I really see this as a path you’re going in or the first step is kind generic tagging of images, right? So what’s in an image? Are their people in it? What are the emotions? Stuff like that that’s pretty generic. And that’s kind of the era we’re in right now where we see a lot of success and where we can really automate these tedious tagging tasks at scale pretty convincingly.

I think the challenge right now is to move to kind of the next step, which is to personalize these tags. So, basically provide tags that are relevant not just to anyone but to your particular company. So, if you’re a car manufacturer and you want to be able to classify different car models. If you’re a retailer, you may want to be able to do fine grain classification of different products. So that’s the big challenge I see now and that’s definitely where we are headed and where we’re focusing on in all apps.

Henrik:  And, as of November 2016, how do you see image recognition changing?

Jonas:  Well, really where I see it changing is, as I said, it’s going to be more specific to the individual customer’s assets. It’s going to be able to learn from your guidance. So, basically, how it works now is that you have a large repository of already-tagged images, then you train networks to do classification. What’s going to happen is that we’re going to add a piece that makes this much more personalized, much more relevant to you, and where the system learns from your existing metadata and your guidance, basically, as you curate the proposed tags.

Another thing I see is video, it’s going to be more important. And video has that temporal component, which makes segmentation important, and that’s how that differs from images. So there’s that, and also the much larger scale that we’re looking at in terms of processing and storage when we’re talking about video. Basically, video is just a series of images, so when we develop technologies to handle images, those can be transferred to the video pieces, as well.

Henrik:  Jonas, what advice would you like to share with people looking at image recognition?

Jonas:  Well, I would say start using it. start doing small POCs [proof of concepts] to get a sense of how well it works for your use case and kind of define small challenges that, small successes you want to achieve and just get into it. This is something that is evolving really fast these days, so getting in and seeing how it performs now, then you’ll be able to provide valuable feedback to companies like Adobe. So you can basically impact the direction that this is going in. It’s something we value a lot. It’s really valuable to us that when we run beta programs, for instance, that people come to us and say, “You know, this is where this worked really well. These are the concrete examples where it didn’t work that well,” or, “These are specific use cases that we really wish that this technology could solve for us.”

So now is a really good time to get in there and see how well it works. And also, I’d say, just stay on top of it. Stay in touch because, as I said, this evolves so fast that you may try it today and then a year from now things can look completely different, and things can have improved tremendously.

So that’s my advice. Now is a good time. I think the technologies have matured enough that you can get real solid value out of them. So this is a good time to see what can these technologies do for you.

Henrik:  Jonas, where can we find more information?

Jonas:  Yeah, so we just at Adobe launched what we call Adobe Sensei, which is the collection of all the AI and machine learning efforts we have at Adobe. And going, just Googling that, and going to that website, that will be updated with all the exciting things that we are doing in that space. And I would recommend that you keep an eye on that because that’s something that’s going to really evolve the next few years.

Henrik:  Great. Well, thanks, Jonas.

Jonas:  Yeah, you’re welcome.

Henrik:  For more on this, visit Tagging.tech.

Thanks again.


 

For a book about this, visit keywordingnow.com


Leave a comment

Tagging.tech interview with Nikolai Buwalda

Tagging.tech presents an audio interview with Nikolai Buwalda about image recognition

 

Listen and subscribe to Tagging.tech on Apple PodcastsAudioBoom, CastBox, Google Play, RadioPublic or TuneIn.

 

Transcript:

 

Henrik de Gyor:  This is Tagging.tech. I’m Henrik de Gyor. Today, I’m speaking with Nikolai Buwalda. Nikolai, who are you, and what do you do?

Nikolai Buwalda:  I support organizations with product strategy, and I’ve being doing that for the last 15 years. My primary focus is products that have social networking components, and whenever you have social networking and user‑generated content, there is a lot of content moderation that’s a part of that workflow.

Recently, I’ve been working with a French company, who’s launched their large social network in Europe, and as a part of that, we’ve spun up a startup that I’m the Founder of called moderatecontent.com, uses artificial intelligence to handle some of the edge cases when moderating content.

Henrik:  Nikolai, what are the biggest challenges and successes you’ve seen with image recognition?

Nikolai:  2015 was really an amazing year with image recognition. A lot of forces really came to maturity and so you’ve seen a lot of organizations deploy products and feature sets in the cloud that used or depend heavily on image recognition. It probably started about 20 years ago with experiments using neural networks.

In 2012, a team from the University of Toronto came forward with a real radical development in how neural networks are used for image recognition. Based on that, there was quite a few open source projects, a lot of video card makers also developed hardware that supported it, and in 2014 you saw another big leap by Google in image recognition.

Those products really matured in 2015, and that’s really allowed for a lot of enterprises to have a very cost effective ability now to integrate image recognition into the work that they do. So 2015 really has seen, in the $1000 range, the ability to buy a video card, use an open source platform, and very quickly have image recognition technology available to your workflow.

In terms of challenges, I continue to see two of the very same challenges existing in the industry. One is the risk to a company’s brand, and that still continues.

Even though image recognition is widely accepted as a technology that can surpass humans in a lot of cases for detecting patterns and understanding content, when you go back to your legal and to your privacy departments, they still want to have an element of humans reviewing content in the process.

It really helps them with their audit, and their ability to represent the organization when an incident does occur. Despite companies like Google going with an image recognition first passing the Turing test, you still end up with these parts of the organization who want human review.

I think it’s still another five years before these groups are going to be swayed to have an artificial intelligence machine‑learning first approach.

The second major issue is context. Machine learning or image recognition is really great at matching patterns in content and understanding these are all the different elements that make up some content, but they are not great at understanding the context ‑‑ the metadata that goes along with a piece of content ‑‑ and making assumptions about how all the elements work together.

To illustrate this, it’s probably a very good use case that’s commonly talked about, which is having a person pouring a glass of wine. Now, in all kinds of different contexts, this content could be recognized as something that you don’t want associated with your brand versus not being an issue at all.

If you think about somebody pouring a glass of wine, say at a cafe in France versus somebody pouring a glass of wine in Saudi Arabia. Between the two, there’s very different context there, but very difficult for machine to draw conclusion about the appropriateness of that.

Another very common edge case that people like to use as example is the bicycle example where machines are great at detecting bicycles. They can do amazing things, far surpass the ability of people to detect this type of object, but if that bicycle was a few seconds away from being into some sort of accident, machines are very difficult at detecting this.

That’s where human review ‑‑ human escalations comes into play for these types of issues and still represent a large portion of the workflow and the cost in moderating content. So, mitigating risk within your organization to have some sort of person review of content.

Then to also really understand the context are two things that I think, in the next five years, will be solved by artificial intelligence and will really put these challenges for image recognition behind them.

Henrik:  As of March 2016, how much of image recognition is completed by people versus machines?

Nikolai:  This is a natural stat to ask about, but I think, with all the advancements in 2015, I really like to talk about a different stat. Right now, anybody developing a platform that has user‑generated content has gone with Computer Vision Machine learning approach first.

They’ll have a 100 percent of their content initially reviewed with this technology and then, depending on the use case and the risk profile, a certain percentage gets flagged and moved on to a human workflow. I really like to think about it in terms of, “What is the number of people globally working in the industry?”

We know today that about 100,000 to 200,000 people worldwide are working at terminals moderating content. That’s a pretty large cost and a pretty staggering human cost. We know these jobs are quite stressful. We know they have high turnover and have long‑term effects on the people doing these jobs.

The stat I like to think about is, “How do we reduce the number of people who have to do this and move that task over to computers?” We also know that it’s about a thousand times less expensive to use a computer to moderate this. It’s about a tenth of a cent per piece of content versus about 10 cents per content to have a piece of content reviewed with human escalation.

In terms of really understanding how far we’ve advanced, I think the best metric to keep is how we can reduce the number of people who are involved in manual reconciliation.

Henrik:  Nikolai, what advice would you like to share with people looking into image recognition?

Nikolai:  My advice is, and it’s something that people have probably heard quite a bit, which is it’s really important to understand your requirements and to gain consensus within your organization about the business function you want image recognition to do.

It’s great to get excited about the technology and to see where the business function can help, but it’s the edge cases that can really hurt your organization. You have to gather all the requirements around.

That means meeting with legal, privacy, security and understanding the use case that you want to use image recognition for and then the edge cases that may pose some risks to your organization. You really have to think about all the different feature sets that go into making a project really successful with image recognition.

Things that are important is how it integrates with your existing content management system. A lot of image recognition platforms use third parties, and they can be offshore in countries like the Philippines and India. Understanding your requirements for sending content over there, your infosec department is really important to know how that integrates.

Having escalation and approval workflows, this is really going to protect you in these edge cases where there is the need for human review. That needs to be quite seamless as there’s still a significant amount of content that gets moderated and approved this way.

Having language and cultural support, global companies really have to consider the impact culturally of content from one region versus another. Having features and an understanding built into your image recognition that it can adapt to that is very important.

Crisis management, this is something that all the big social platforms have playbooks ready to go for. It’s very important because, even if it’s, like I said, one image in a million that gets classified poorly, it can have a dramatic impact in media or even legally for you. You want to be able to get ahead of it very quickly.

A lot of third parties provide these types of playbooks, and it’s a feature set that they offer along with their resources. The usual feature set you have to think about ‑‑ language filters, image, video, chat protection. Edge case that has a lot of business rules associated with is the protection of children, social‑media filtering.

You might want to have a wider band of guardrails to protect you on response rate and throughput. A lot of services have different types of offerings. Some will moderate content over 72 hours, and others you need response rates within the minute.

Understanding your throughput and response rate that’s required is very important and really impacts the cost of the offering that you are looking to provide. Third‑party list support ‑‑ a lot of companies will provide business rule guidance and support on the different rule sets that apply to different regions around the world.

That’s important to understand which ones you need and how to support it within your business process. Important to demonstrate control of your content is having user flags. Being able to have the people who are consuming your content, the ability to flag content into workflow to work through that demonstrates one of the controls that you need to often have in place and the edge cases.

The edge cases are where media and legal really has a lot of traction and are looking for companies to provide really good controls for protecting themselves. Things like suicide prevention, bullying, and hate speech can really dramatically…just one case can have a significant impact on your brand.

The last item is a lot of organizations for a lot of different reasons have their content moderation done within their own organization. They have the human review within their own organization and so having training of that staff for some of the stressful portions of that job and training for HR is very important. It is something to consider when building out of these workflows.

Henrik:  Nikolai, where can we find more information about image recognition?

Nikolai:  The leading research for image recognition really starts at the ImageNet competition that’s hosted at Stanford. If you Google ImageNet in Stanford, you’ll find that the URL isn’t that great and officially it’s called the ImageNet Large Scale Visual Recognition Challenge. This is where all the top organizations, all the top research teams in image recognition compete to have the best algorithms, the best tools, and the best techniques.

This is where all the breakthroughs in 2012, 2014 happened. Right now, Google is the leader, but it’s very close and image recognition at that competition is certainly at a level where these teams are far exceeding the capability of humans. So from there, you get to see all the tools and techniques that the latest organizations are using, and what’s amazing is the same tools and techniques they use on their platforms that exist for integrating within your own organization.

On top of that, the competition between video card providers, between AMD and NVIDIA, has really made the hardware to support this to allow for real‑time image recognition at a very cost-effective manner. The tools that they talk about at this competition leverage that hardware and so it’s a great starting place to understand what the latest techniques are and how you might implement them within your own organization.

Another great site is opencv.org or open computer vision, and they have taken a built‑up framework around taking all the latest tools and techniques and algorithms and packaging them up in a really easy‑to‑deploy toolset. It’s has been around for a long time and so they really have a lot of examples, a lot of the background about how to implement these types of techniques.

If you are hoping to get an experiment going very quickly, using some of the open source platforms from ImageNet competitions and using OpenCV together you can really get something up very quickly.

On top of that, when you’re building out these types of workflows, you need to work closely with a lot of the nonprofits that have great guidance on what are the rule sets, what are the guardrails you need to have in place to protect your users and to protect your organization.

The Facebook has really been a leader in this area and they have spun up a bunch of different organizations they work with ‑‑ the National Cyber Security Alliance, Childnet International, connectsafely.org ‑‑ and there are a lot of region‑specific organizations that you can work with. I definitely recommend that using their guardrails will really be a great starting point for a framework when understanding how image recognition can moderate your content, how image recognition can be used in ethical and legal manner.

In terms of content moderation, it’s a very crowded space right now. Some of the big partners, they don’t talk a lot about their statistics, but they are doing a very large volume of moderation. Companies like WebPurify, Crisp Thinking, and crowdsource.com, they all have an element of machine learning and computer and human interaction.

The cloud platforms like AWS and Azure have offerings for the machine learning side. Adobe definitely is a content management platform. They have great integrated software package if you use that platform.

Another aspect, which is quite important, is a lot of companies do their content moderation internally, and so having training for that staff and training for your HR department is very important. But all in all, there are a lot of resources, a lot of open source platforms that make it really easy to get started.

TensorFlow, which is an open source project from Google, they use it across their platform. I think they have…The last I checked, it was about 40 different product offerings that use the TensorFlow platform, and it is a neural network based image recognition type technology. It’s very visual and it’s very easy to understand and can really help reduce the amount of time to go to production with some of this technology.

Other open source projects, if you don’t want to be attached to Google, include CaffeTorchTheano and NVIDIA. They have a great offering tied to their technology.

Henrik:  Well, thanks Nikolai.

Nikolai:  Thank you, Henrik. I’m excited about content moderation. It’s a topic that’s not really talked a lot about, but it’s really important and I think in the next five years we are really going to see the computer side of content moderation and image recognition take over, understand the context of these items, and really reduce the dependency on people to do this type of work.

Henrik: For more on this, visit Tagging.tech. Thanks again.


 

For a book about this, visit keywordingnow.com