With the advent of digital, pro photographers have been taking more and more pictures. The downside is that they end up spending more and more time in post-production, having to search within thousands of images for the winning photographs. This is what Hungary based startup Picturesqe solves: using a mix of machine learning and A.I, their solution cuts the editing time by 80%, honing rapidly and efficiently into the top candidates. We sat them with them to learn more:
– A little about you, what is your background?
The founders, Daniel, and Peter are visionary entrepreneurs. They have their own prospering B2B and B2C software companies with many successful projects over the last decade. They are also advanced photographers too, they have a first-hand experience what a painful process photo selection can be. Their vision is to use cutting edge computer vision and AI solutions to turn this into a joyful task.
Our lead researcher – Szabolcs – is a physicist. He spent three years in the Santa Fé Institute, in New Mexico as a PhD fellow, and has attended many AI projects since then. Attila – our lead developer – spent years with digital image processing and developing software that analyzes facial micro-expressions digitally. Marton has an extensive background in market research and marketing. He previously founded a marketing startup.
– Explain Picturesqe. What does it solve and how does it work?
Picturesqe is a client application. It has smart tools and handy user interfaces to save time and money for professional photographers. Picturesqe uses advanced computer vision algorithms and AI, which automate some parts of the photo selection process to save 80% of the time spent on browsing and selecting photos. We analyze the photos offline after you have copied them onto your computer; we compress the digital information from each photo and upload a small, but rich image descriptor to our cloud where the intelligent algorithms run. Since the computing capacity in the cloud is flexible this is a really quick process. To sum it up, photographers don’t have to upload their pictures and still can enjoy the features provided by an AI. Since most of the heavy calculations are done in the cloud, users don’t need a “power-station” desktop.
We analyze the photos offline after you have copied them onto your computer; we compress the digital information from each photo and upload a small, but rich image descriptor to our cloud where the intelligent algorithm runs. Since the computing capacity in the cloud is flexible, this is a really quick process. To sum it up, photographers don’t have to upload their pictures and still can enjoy the features provided by an AI. Since most of the heavy calculations are done in the cloud, users don’t need a “power-station” desktop.
– Where does Picturesqe work? Is it a mobile app, a SaaS, a lightroom plugin?
We offer Picturesqe now as a SaaS solution. It has a local client that communicates with our servers in the cloud. We use a computing cloud to do the calculation heavy tasks and send the results back to the client. Either a desktop or a mobile (iOS) client can be used.
– What are the criteria your engine uses to define what a good picture is?
Photo aesthetic is a very subjective field; we don’t believe machines will ever learn what a good picture is in the art sense. But photo quality is measurable to some extent; there is certainly is a part of it which is objective. Sharpness of the area of interest, exposure, composition, color balance, etc. – there are obvious or hidden rules for these measures even if they are too complex to be described by some simple decisions – this is where machine learning comes in. We subtract the interesting features from a large set of manually ranked images and learn the hidden rules using ML algorithms. This is how the criteria for a good quality image arise.
The definition of a photo group in Picturesqe is that you would want to keep only one photo out of the set. So we are collecting images taken close in time, space and with high visual similarity in the visually attended regions. Within such a group it is easy to compare the photos with Picturesqe’s smart tools and the photographers can make the final, subjective decision which photo to keep or delete.
– Are you using machine learning and AI?
Machine learning is a subfield of AI. Picturesqe is an AI application with machine learning. We use intelligence learned from talking to photographers and intelligence learned from examples.
– A good news photo doesn’t have the same aspect as a good landscape image or family pictures. Does Picturesqe make the difference?
Digital still images are pretty complex creatures. There are different levels of abstractions in them – starting from a set of colored pixels to a living story. Currently, we do not do semantical photo understanding. Images within different photographic categories, like portrait, landscape, news, sports, e.t.c. have few things in common on the visual level (sport photos are taken usually with a larger aperture, faster shutter speed, while landscapes usually operate with high depth of field and longer exposition), so it’s possible to separate them out. We understand patterns and will make groups according to similarities. To help in the photo selection process as much as we can, we don’t have to understand the stories behind the pictures, …yet. Picturesqe is not a photo categorization software right now, but we’d like to add such capabilities soon.
– How about cultural difference? What is considered a great picture in China for example, might not meet the same response in the UK?
Our system is able to learn from each photographer separately after there are
enough sample photos collected on his account. His/her preferences will then appear in the automatic ranking. There are different photo making styles which we can cater to.
– Aren’t you worried that after a while all the images will look the same?
We are not afraid that the AI used here will make all the photos look the same after a while since we only want to touch the objective part of photo quality. It doesn’t make decisions, just suggestions. We mark images as good or bad quality, but let the users make the final decisions to delete them or to keep them for post-production.
– Who is Picturesqe’s typical user?
Any photographer that has a lot of pictures – event or fashion photographers, or serious enthusiasts. Professionals who take tremendous amounts of pictures and would like to have a tool that can save their time and therefore, money. Also, any enthusiasts who constantly worry about selecting their best shots.
– What would you like to see Picturesqe offer that technology cannot yet deliver?
We would like this platform to be able to develop an AI that can do what a professional photographer would do manually with our pictures – just fully automated and keeping in mind our individual preferences. Such platform would be invaluable to any business that processes a huge amount of digital pictures.
Author: Paul Melcher
Paul Melcher is the founder of Kaptur and Managing Director of Melcher System, a consultancy for visual technology firms. He is an entrepreneur, advisor, and consultant with a rich background in visual tech, content licensing, business strategy, and technology with more than 20 years experience in developing world-renowned photo-based companies with already two successful exits.