Editors note: For the latest on image recognition and auto-tagging, read Do DAM Professionals Think Image Recognition is Finally Ready for Digital Asset Management? and Your Auto-tagging Recipe for Success.
As of today, human intelligence is still king when it comes to accuracy and efficiency with your metadata.
But just because image recognition and AI aren’t 100% ready to tackle your metadata problems, doesn’t mean they won’t be one day. In fact, when they are it will finally be time to automate metadata tagging — which will free up your DAMster’s time for other tasks. Until that time, you’ll need to call upon the expertise of a dedicated DAM admin.
One of the major goals of a DAM system is to make it easy for users to search for and find the assets they need — and the success of that search depends on accurate and consistent metadata. Today, some DAM vendors insist that image recognition can eliminate the burdens of entering metadata. However, we’ve found evidence to the contrary that shows automated tools produce inaccurate and inconsistent metadata, which can frustrate users during their search process — and create more metadata work in the long run.
So today, we’re going to explore:
- What all the hype is about image recognition
- How accurate is artificial intelligence in tagging visual content?
- Some of our findings concerning AI machine learning and image recognition
- Defining what level of metadata accuracy is accurate enough
- Our final recommendations
Why all the hype about image recognition?
Image recognition has become a hot topic because the number of digital assets, and the number of people searching for those assets, has dramatically increased over the last five years. In 2011, our largest data set had 250,000 assets. Now, we see customers with asset counts in the millions, and they’re growing rapidly. It’s no surprise the biggest challenge we hear from customers is that they don’t have time to enter metadata.
In these massive digital asset management libraries, users can struggle to find what they need if assets are not clearly tagged with accurate metadata. Our onsite field studies have shown that users will spend around two minutes searching for an asset before they give up and turn to stock photography houses or images saved on their own desktops.
With properly tagged assets, coupled with a team that understands what the metadata parameters are, this same task takes only seconds. But just because you understand the importance of metadata, doesn’t mean your team always has the time to apply the metadata to all your assets — which is why artificial intelligence (AI) seems like such a great solution.
In fact, many DAM vendors have partnered with image recognition services to automatically tag metadata on their visual content. If you’re looking at DAM systems, buzzwords like “AI,” “automation,” and “machine learning” are catchy. But are they helpful or just hype at this point?
Is Artificial Intelligence helpful in the “heavy lifting” of tagging files? Don’t believe the hype (at least not yet)!
Although we can’t tell you the reliability of each vendor’s AI implementation, we can share our research into the same open source auto tagging services they’re using. We’ve been actively exploring open source libraries and services for the past five years (if you’re curious, we’ve tested Imagga, Clarifi, CloudSight, Imprezzeo, and Google API, to name a few).
We’ve seen the accuracy of auto-tagging services increase over the years — and it may continue to do so — but it’s still working towards a point of being consistently useful.
What we found with AI machine learning
Machine learning can be very powerful when there’s a large enough data set. The current challenge is the data set required to train the service needs to be larger than the amount of data available to any single customer. Since product imagery is unique from one brand to the next, each brand would need its own data library. For now, this isn’t a cost-effective solution for our customers.
What we found with image recognition
Our team connected various image recognition services to the Widen Collective via the APIs of these other tools and code libraries. The services returned lists of general metadata tags and confidence values. Although results were usually accurate, there were still incorrect tags despite having high confidence values. Some services were better at identifying faces. Other services were better at identifying objects. But no service excelled at both.
That said, it’s obvious the applications are on the right track. In fact, our research suggests that image recognition services tag the right subject about 80 percent of the time.
Define what “accuracy” means for your teams
Automated tags generated with up to 80% accuracy is a good start. Theoretically, that means 80% of your files required less manual keyword tagging. But there’s a difference between accurately identifying what’s in a photograph and accurately identifying the terms users search to find that file.
Even if the keywords are correct, if assets aren’t tagged with metadata that your users will enter, no one will find and use the assets.
Here is an example of a customer image from Briggs & Stratton. Their goal was to have generic tags automatically applied to assets to supplement the product specific metadata that they already applied.
All of these tools return accurate tags — “lawn mower,” “grass,” “outdoors,” “man,” and “wheel.” The third tool, Cloudsight, specializes in physical products. It has the most accurate tags of “black and red push mower,” “men’s gray and blue denim jeans,” and “dress shirts.”
Although these tags are accurate, will your users — dealers, partners, web designers, sales teams, etc, use these terms to search for the files they need? Do you expect a web designer to search “outdoors” or “men” or “dress shirt?” More likely they will be looking for a “lifestyle” shot. In the the case of “black and red mower,” for a manufacturer of lawn mower engines and brand colors of red and black, the number of search results returned with these tags will be large and it is likely for a user to abandon their search rather than page through hundreds of results.
Fortunately, you can determine how your teams and partners search for files. If you have a DAM system, your analytics should provide search query information. You can also use questionnaires and interviews to find out how your users search.
As you can see it’s really all about having the metadata act as “the great connector” between what someone is actually searching for, with assets that precisely fit the actual search criteria (which has to be determined, agreed upon and implemented by actual people). So even in the case of 80% accuracy, the specificity percentage can often be much lower.
The other 20%
If image recognition is 80% accurate, what happens with the 20% of inaccurate metadata tags?
In the previous example of the Briggs & Stratton mower image, there are several incorrect tags such as values of “motorcycle,” “accident,” “family,” “bike,” and “transportation system.” Will your busy teams be diligent enough to correct the inaccurate metadata tags? If not, your assets now have bad metadata, which is worse than no metadata at all. At least with no metadata, a digital librarian can track down assets and tag them. Bad metadata means the assets will never be found and are just taking up storage space.
What we recommend ...
Through our research and conversations with customers, we found that image recognition tools will offer high value once they are able to accurately and consistently apply tags. Until then, they’re simply not dependable enough to use with confidence.
If you’re considering a DAM system touting AI and image recognition, keep your use cases and metadata strategy top of mind. Does the software properly tag your assets? Can it identify the nuances between your various products and parts?
A proven solution to the metadata challenge
Remember, your goal is to be efficient and get the most value out of your assets for a better ROI. If assets are getting lost in the shuffle, and team members are downloading stock photos instead of using approved imagery, there’s a good chance you’re losing a lot more than you’ve hypothetically gained by automating your metadata.
Investing in a DAM admin (digital librarian), a true visual content superhero, can go a long way in maximizing your assets, and helping your organization grow. In fact, we’ve found that companies with a dedicated DAM admin use their assets twice as many times as organizations without a dedicated admin.
Want to continue the DAM conversation?
Get in touch with one of our advisors.