Original article can be found here (source): Deep Learning on Medium
Our AI Editor is Almost Live
We’re Saving the Best for Last
Hi all! First off, you should be reading this from home. It is crucial to stay safe from the current pandemic by social distancing. Anyway, at Homedeck we’ve been shifting towards becoming more active with our research. And in case you aren’t familiar, we provide a fully automated photo editing service for real estate photography. Our service is powered by artificial intelligence, hence allowing us to be fully automated—exactly zero human interaction.
As we tinker with the hooks and knobs, we want to provide folks with some insight on just how it all works. This will allow photographers to best take advantage of our service, in turn allowing us to iterate most productively. The Homedeck Beta is open for use and completely free. There are no user accounts; no credit cards; no questions asked. Our hope is that we can begin to prepare photographers for an entirely new workflow, one where editing simply ceases to be a task. So let’s get right to it.
The Magic Behind the Edit Button
Our editing solution is built with two goals in mind: quality and speed. From the outset, Chris and I have aimed to completely automate the process of editing real estate photos because we think it’s a chore. The vast majority of photographers and editors use the same tools, the same editing filters, and the same metrics for subjective quality. So if we could look at a bunch of images and say which ones looked good enough to deliver to a client, then why couldn’t we write an algorithm to do the same? And of course if we did, then we would be able to free up tons of time to either do more shoots or something more enjoyable.
Much like human editors, Homedeck AI edits images in two steps: fusion and enhancement. When you upload a set of photos, an available AI editor will claim the shoot and download all the images you just uploaded. It then breaks them up into groups, based on the number of brackets you specify, and then merges them using a fusion algorithm we call Deep Fusion.
Deep Fusion, like all else, is inspired directly by manual blending that many human editors do. In crafting Deep Fusion, we wanted something that was expressive enough to capture more advanced blending methods like flambient (flash+ambient). More elementary fusion methods like HDR and Exposure Fusion simply fail to capture some of the more semantically-important parts of the exposures.