Viesus™ at Visual1st 2020: Photos as they’re meant to be, supported by Machine Learning
Today we explained at Visual1st 2020 Virtual Event how we apply machine learning to further improve the performance of Viesus. Machine learning helps us to achieve a better understanding of what’s happening in a photo and to leverage this information to do a better enhancement. We are very excited about how machine learning is helping us to improve even further upon our proven image enhancement quality. The video below demonstrates in a nutshell how we do the analyzing of a photo. The analysis is input for the different enhancement steps.
During the panel discussion at Visual1st we discussed the following questions:
Your main markets for Viesus?
For Viesus it was first photo & printing labs, but over the last years we started to help other businesses as well. For example, companies like second hand car sellers, travel apps and even online flower shops, all benefit from better images of what they are selling (you don’t want murky photos of your products). This expansion was possible because we made Viesus available for multiple platforms, such as Android, iOS, macOS.
What part of your solution is based on machine learning?
Machine learning is part of the initial analysis that we call image understanding. Before taking action, Viesus analyses the image, with both traditional image analysis tools, and machine learning: what’s on the photo, how are the colors, how is the contrast, where are people and faces, are there any red-eyes, how are color distributions, etc. With this understanding, Viesus can then perform a tailored and custom enhancement, exactly for what specific image needs.
What kinds of training sets does your solution leverage?
We have an image database built up over the last 20 years. This database contains tens of thousand of labeled images, and we leverage it to train Viesus for specific deep learning tasks. Obviously some tasks require specific labeling. In this case we either apply the labeling on our dataset, or we use third parties labeled data.
As you’re really pushing the envelope to cater to new use cases or cater to them in a different way (more “lights-out” automatic, using privacy-sensitive user data etc.), have you encountered pushbacks from (prospective) clients? And how have you responded?
In some very rare cases VIESUS can encounter a peculiar photo that it can not enhance properly yet. When we receive this type of feedback from a customer we improve VIESUS so that it can handle the photo correctly. We first identify how can we recognize the problematic situation on the photo. And then, together with our color scientist we define an enhancement that satisfies customers’ expectations. We have been operating in this way for a while, so these cases are already rare.
We are used to dealing with privacy sensitive data. Data are used in a secured sandbox and thrown away after the feedback is solved.
What do you see your technology could possibly do 5 years from now?
In 5 years from now we are applying much more AI for image enhancement to make photos as real as possible. We are now pushing the boundaries of the triangle speed, quality and costs to apply AI. In 5 years time we will have affordable dedicated AI hardware to overcome limitations from speed and costs. At the moment we can achieve with quality much more than we ship today. Because the processing power is too expensive, or it costs too much process time for industrial scale. We are confident this is solved in the near future.
What are your barriers to entry: How much of a barrier to entry are your training sets? I.e. are they hard to come by/develop for others who want to enter the market, e.g. comparable to how social solutions have network effects that are hard to break into for newcomers? Or are your client relations the barrier to entry? Or your channel partners? Or your AI team? Or…?
For us it actually lowers barriers, we can solve feedback much faster then before. By recognizing an issue in a photo with Machine Learning. Testing the enhancement result is more a barrier which costs us more time, we do not use AI to determine if the result of enhancement is good or bad. To determine if a photo is good can only be done by the human eye for now. To reach a common agreement on what the human eye thinks a good photo is we need to do a blind test with a group of persons.
How did you make your decision to go B2B vs. B2C or both?
We focus on B2B because Viesus is built for industrial scale, to fit right into a customer’s workflow, software solution or app. With new developments coming up we might go more into the B2C market when we can uphold our values of speed, costs and quality. Get in touch with us and stay connected to get latest information on this.
When you think about the variety of vendors in our photo & video ecosystem, are there particular types of vendors you see great partnering opportunities with?
About the variety, we are excited to see other image enhancers and other imaging companies in the ecosystem addressing areas we do not cover, as we do not do beautification but “real”-ization. Everyone is pushing the limits with new technology so these are very exciting times. As mentioned earlier, everyone who does something with photo (photo centric) should do a slight touch of enhancement, that makes a lot of potential partners. Do not miss the opportunity to start easy with photo enhancement in your solution or company and reach out.
Go to www.viesus.com for more information.