Over the past six or seven years, I’ve talked to many DAM practitioners, consultants and developers about using AI for tagging assets. It would be fair to say that their experiences have varied widely, and most of the time it has taken significant effort before auto-tagging came close to fulfilling its promise. A surprisingly high proportion had trialled AI tagging for a short period, ended the project, and were waiting for the industry to mature further before trying again.
The most common issue raised was that metadata provided by the AI engine was too generic and couldn’t meet domain-specific requirements. For a car manufacturer, it’s not enough for an image to be tagged “car”: to be useful, the AI needs to be able to determine type, brand, and even model and year. Another concern was regional differences in vocabulary, such as between US and UK English. The footwear industry is particularly impacted by this issue, with terms like “sneakers” and “pumps” having quite different meanings in different countries (or even within a country).
Considerations to keep in mind when evaluating AI services
But many projects only evaluate a single AI provider. Feedback from customers and industry contacts, and tests we’ve performed here at OpenText TM, have highlighted the variation in metadata coming from different AI services when analyzing the same images. Every service has different strengths and will provide a unique perspective on the content. For example, in a photograph of a city street featuring someone wearing a red dress, different AI services might be better at:
- Identifying people in the picture, either from a library of celebrities, or as a person appearing in other pictures in your DAM
- Identifying the color and style of the dress, or linking the asset to a product by comparing it with reference images
- Identifying the location of the photograph using landmarks
- Extracting text from the image such as street signs, license plates or store names
- Identifying brand logos, objects, animals, plants, faces, ages of people, or other elements
- Providing crop suggestions based on points of interest within the image
- Compare images of the location taken over time and track corrosion, vandalism, or plant growth
- Highlight anti-social, illegal and sexual behavior such as drug use or nudity
- Any other area of image analysis
Each organization is looking for something different from AI, but no single service does everything well. Finding the right AI partner, whether one of the better known from Azure, Google or AWS, or a specialized provider, is vital for getting real value out of the technology.
However, some DAM providers exclusively use one AI service, either an in-house engine or a partner’s. That results in DAM users being locked in with a single AI partner, unable to compare different services to find the best for their specific use case or use cases. If it’s specialized in identifying people, an entertainment news service would find it valuable, but it is of limited use for a company selling building supplies.
Easily integrate with the AI provider of your choice
The approach we have taken at OpenText is to preserve customer choice when selecting an AI image analysis service. OpenText Media Management TM integrates with Azure, Google and AWS out of the box, but the integration can be easily extended to connect to any AI provider. We believe the best approach is to empower our customers to select the best AI partner (or partners) for their needs. AI auto-tagging is a potentially powerful assistant for DAM users, but it isn’t a “one size fits all” situation. As with all technology decisions, selecting the right partner is vital for success.
If you’re interested in hearing more from us on this subject, you can come and find us at DAM Europe 2022, which is taking place from 22-23 June.