Quick selection rule: If I can see it, shouldn't Pixelmator see it? (And a Show Selection Preview "Rewind" request...)

Talk about Pixelmator Pro, share tips & tricks, tutorials, and other resources.
User avatar

2022-01-22 18:22:18

Hello and thanks for such an extraordinary product. I only wish I'd found it years earlier! Pixelmator Pro is a cornerstone of some of the digital work I do. I am grateful for the incredible effort you team has put into the the brilliant original product and in your striving to keep improving it.

I appear to be having an issue with the Quick Selection tool. While the actual use of it is as easy as point and click, sometimes it appears to be having difficulty distinguishing boundaries. This of course matters because it is an essential aspect of the tool. In my version of Pixelmator Pro (Abracadabra, v2.3.5), sometimes it has a pixel-level accuracy and sometimes it doesn't.

I sometimes work with extremely fine and difficult boundaries, so not finding a crisp boundary can be the crucial difference in an alteration--although, admittedly, sometimes a lumpy boundary can work. I have come to a simple but perhaps difficult-to-enact axiom: If I can see it, Pixelmator Pro should be able to see it.

Sometimes you can clean-up a Quick Selection by changing the Add or Subtract options after a first pass attempt. For instance, if one were cutting out a pair of jeans against a similar-colored background, it could get lumpy. In many instances, one can then switch to the Subtract option, click outside the main cut-out target and the fine boundary can be found. One can thus work from the Inside-->Out, then Outside-->In against the selection. (Given there is image detection in the algorithms, perhaps this is already another place where it is used under-the-hood?)

I know this can get difficult at the pixel level. Is there a "slider tool" sub-menu accessible to Quick Selection that might allow me some variation? One imagines that if one could run one's cursor along the naked-eye boundary in what one hopes is a foregrounded selection, one might clean things up? Yet in trying to do so, sometimes lumpiness increases; other times the boundary suddenly jumps farther inside the target area, or appears lost such that it suddenly grabs an entire part of the background area.

In a side note, I've found that the Quick Selection-Show Preview option can often be helpful; yet if one has a terribly fast processor (M1!) these Previews become "Lightning" Previews, and one might only realize that the selection one wants was 2 versions "back"; one consciously recognizes the selection one wants _after_ it has been lost as the Lightning Preview moves on to the next "bolt." That is, one's human ability lags behind the computing ability to generate Show Preview opportunities; the capture one liked but has been lost in the cursor-moved selections cannot then easily be found again. To b clear, one cannot sometimes place one's cursor precisely enough to revive a missed-opportunity Preview selection. One might prefer that these "bolts" be "recorded" so that the advantages/disadvantages in one's Preview opportunities can be reviewed/rewinded in a slower-than-real-time motion in such a way. One might right-click to "expose" a Preview-opportunity frame series, such that, brought to the forefront, one can then page through them...

Back to the main concern: I have submitted a corner-case example, a clear plastic halo shot against a blue sky. One can see how QS gets close in the original image (001), but remains lumpy as can be seen in the selection re-pastes (002, 003)...

Thanks for any suggestions!

Image
Image
Image
User avatar

2022-01-22 18:42:11

Maybe you could upload that portion of the original image.
User avatar

2022-01-22 19:58:55

I have found selecting first with the Quick Selection Tool, followed by refining with the Free Selection Tool to be helpful for addressing the ragged edges you mentioned.
User avatar

2022-01-26 15:19:48

Thanks for the feedback, JM Vore. The Quick Selection uses a machine learning-powered algorithm to make selections, so it essentially learns from the examples you supply to it. A sample image would, indeed, help in this case. We could then try to include it in our training dataset and see how that works. :pray: Also, we'll keep the selection preview speed in mind. Perhaps there's something we can improve there.

For now, I'd probably suggest using the Select and Mask tool to refine your selections. There are options like Roundness you can use there to soften up the ragged edges, brush tools you can use to add or remove areas from a selection by simply painting over them, and more.

As for the idea that Pixelmator Pro should "see" images the way we, humans, do — I don't think I can agree. Or, at least, I don't think we're there yet technologically. Let's say you give 10 people the same image and tell them to select an object in it. You'd get 10 close but different selections. A computer algorithm, in the meantime, will select the same object the same way every single time (see Pixelmator Pro's Select Subject tool, for instance). That's because a lot of what we see is an interpretation, and we often fill the gaps or see certain patterns thanks to our real-world experience of how things look or what properties different objects have (like the semi-transparent plastic in your example image). For a computer algorithm, this is an extremely difficult, if at all possible, task.