Adobe demonstrates future Photoshop tool that uses machine learning to select image subjec...

Posted:
in Mac Software
Adobe has teased one of the AI-based features coming to Photoshop in the future, with the 'Select Subject' function simplifying the selection of an object, person, or animal in a scene, using machine learning to determine what is likely to be the main focus of an image.




The video, published on Monday to the official Adobe Photoshop channel, demonstrates Select Subject as a quick and simple way for users to make an initial selection of an element within an image. Compared in the video to existing methods including the quick select tool's edge detection feature, selecting the negative space with the magic wand, or the pen and magnetic lasso tools, Select Subject is shown to highlight the subject in a single click.

The Select Subject tool is also able to work with multiple subjects in a single scene, such as a group of people on a beach. The tool is also available in the Select and Mask workspace, so users can refine the selected area's edge detection, improving the section for more complicated elements like fur on an animal.

For complex scenes that include complicated detail around a subject or an object-filled backdrop, Adobe suggests this isn't an issue, as the tool uses machine learning to work out what visible elements belong to the subject or the background.





Adobe advises that Select Subject should be used as a starting point for selection-related image editing tasks. While the video does show some areas are not detected fully by the tool, like the clothing of one dog walker, it largely appears to correctly select the intended elements, leaving relatively little areas in need of further work.

The Select Subject tool is powered by Adobe Sensei, the firm's AI and machine learning technology used to automate tasks throughout the Adobe Cloud platform. Introduced in November 2016, Sensei has been used to help users find appropriate stock images, powers the Match Font tool, and Photoshop's Face-Aware Liquify operations.

It is unclear when exactly Select Subject will be available to Creative Cloud users, but the video voiceover by Photoshop Product Manager Meredith Payne Stotzner advises it will be implemented in an "upcoming release of Photoshop CC."

Adobe has put considerable effort into its AI-based services, and has previously hinted at a future where more of the tasks could be performed just by machine learning systems. A concept video released in January depicts a tablet-based app that can perform basic edits to an image, including cropping and sharing the result to Facebook, with the app following only "natural language" verbal commands.

Comments

  • Reply 1 of 12
    Being Adobe, this will probably be part of a Photo Add On pack for $9.99/month. 
    kkqd1337dysamoriaMartin57jony0
  • Reply 2 of 12
    Does anyone else find the phrase ‘machine learning’ quite patronising and pretty much meaningless? 

    Such a stupid millennial term 
    dysamoriaaylk
  • Reply 3 of 12
    While the video does show some areas are not detected fully by the tool, like the clothing of one dog walker,

    Or, you know, the model’s left arm in that photo. 
    dysamoriaroundaboutnowSpamSandwichaylk
  • Reply 4 of 12
    dysamoriadysamoria Posts: 3,430member
    kkqd1337 said:
    Does anyone else find the phrase ‘machine learning’ quite patronising and pretty much meaningless? 

    Such a stupid millennial term 
    I came here to say "can we please all stop calling untintelligent softwear algorithms 'artificial intelligence'??"

    There's zero artificial intelligence in existence. All that exists are clever algorithms and slightly human-oriented interfaces. Calling all this crap AI or "machine learning" is utter bollocks. It's absolutely a misuse of the terms.

    However, it has nothing to do with millennials. Please stop hating on a whole generation of people.

    edited November 2017
  • Reply 5 of 12
    MarvinMarvin Posts: 15,322moderator
    dysamoria said:
    kkqd1337 said:
    Does anyone else find the phrase ‘machine learning’ quite patronising and pretty much meaningless? 

    Such a stupid millennial term 
    I came here to say "can we please all stop calling untintelligent softwear algorithms 'artificial intelligence'??"

    There's zero artificial intelligence in existence. All that exists are clever algorithms and slightly human-oriented interfaces. Calling all this crap AI or "machine learning" is utter bollocks. It's absolutely a misuse of the terms.
    I think "machine learning" is an ok term to use. Learning is acquiring knowledge, which is mostly information association. That's what is happening here, the machine is adaptively acquiring associations between shapes and descriptions. Children don't instinctively know animal names, objects, colors etc. They learn by being taught the associations and it's not instant. There are differences between how machines and people start the process, machines aren't spontaneous, but the association process is very similar.

    The terms can imply human qualities like curiosity and spontaneity, which machines don't have and in that sense they are misleading but I think they are close to what they describe. They could use other terms like adaptive semantic mapping but it's just a label to give people an understanding of what's going on and from a marketing point of view, it fits quite well to convey the idea that the computer can identify objects the way that people can.

    The functionality is great and something that has been sorely needed in media. Cropping and masking with even advanced tools is tedious because the computer has had no understanding of the objects in the image, all the pixels were the same. Being able to identify contiguous objects will be a huge benefit for editing and metadata. It will get to a point soon, if not already, where someone will be able to talk to the computer and say things like 'remove my ex from this photo' and it will just do it because it will know who their ex is from facial analysis. To a typical user, that will appear to be intelligent or cognisant of the user's intentions in a way that computers haven't before.
    edited November 2017
  • Reply 6 of 12
    It'll get really interesting when all photos have depth information included. That combined with machine learning could help create pixel perfect masks for still photo retouching and for extremely sophisticated rotoscoping work (possibly on-the-fly).
  • Reply 7 of 12
    dewmedewme Posts: 5,362member
    kkqd1337 said:
    Does anyone else find the phrase ‘machine learning’ quite patronising and pretty much meaningless? 

    Such a stupid millennial term 
    The ML term is used to differentiate adaptive and unbounded systems from fixed ones that have to be preloaded with all potential conditions versus outcomes. But I understand the annoyance level, and I agree that it's also generational and always has been. I'm still annoyed by the gross overuse, misuse, and meaningless use of the word "digital" to imply anything modern or computing related, and with no regard or understanding of where analog systems (like all of life and the entire universe) fit into the digital vs analog picture. Before "digital" it was "solid state" and don't get me started on all things "cyber" or "pinging" outside of the context of an active sonar system. Words suck and English words are the worst.
  • Reply 8 of 12
    dewme said:
    kkqd1337 said:
    Does anyone else find the phrase ‘machine learning’ quite patronising and pretty much meaningless? 

    Such a stupid millennial term 
    The ML term is used to differentiate adaptive and unbounded systems from fixed ones that have to be preloaded with all potential conditions versus outcomes. But I understand the annoyance level, and I agree that it's also generational and always has been. I'm still annoyed by the gross overuse, misuse, and meaningless use of the word "digital" to imply anything modern or computing related, and with no regard or understanding of where analog systems (like all of life and the entire universe) fit into the digital vs analog picture. Before "digital" it was "solid state" and don't get me started on all things "cyber" or "pinging" outside of the context of an active sonar system. Words suck and English words are the worst.
    Are you kidding? The English language is great. It’s continually evolving and expanding, adopting new words from within and outside the culture.
  • Reply 9 of 12
    rufwork said:
    While the video does show some areas are not detected fully by the tool, like the clothing of one dog walker,

    Or, you know, the model’s left arm in that photo. 

    Hey, but it correctly identified the pee-stain under the model!
  • Reply 10 of 12
    polymniapolymnia Posts: 1,080member
    It'll get really interesting when all photos have depth information included. That combined with machine learning could help create pixel perfect masks for still photo retouching and for extremely sophisticated rotoscoping work (possibly on-the-fly).
    I am so ready for this future you describe. My iPhone X is coming this week, I’m excited to begin using depth with today’s capture tools. 

    I wonder what this future means for my expensive Nikon glass. You think they’ll figure out a way to capture depth through existing lenses?
  • Reply 11 of 12
    I think Photoshop users would rather see a significant advancement in the selection tools themselves than the computer self-selecting sloppy silhouettes within a photo.
  • Reply 12 of 12
    I think Photoshop users would rather see a significant advancement in the selection tools themselves than the computer self-selecting sloppy silhouettes within a photo.
    Well... getting at least part of the selection process out of the way using machine learning could be a timesaver. Obviously, additional work would be needed but in the example Adobe showed, it could save a person 5 minutes of work or more. In the case of complex masking jobs it could save considerably more time. I'm really not sure how one will be able to instruct the software exactly what the selection should be... seems like the actual use might be very limited unless using natural language one is able to narrow the target for masking. Suppose you want to select a reflection in a mirror, but not the subject being mirrored?
Sign In or Register to comment.