How to check if an image contains an element I am searchig for

[Chris Angelico]

On Thu, Jun 17, 2021 at 4:43 PM Dan Stromberg <drsalists at gmail.com> wrote:
>
>
>
> On Wed, Jun 16, 2021 at 2:44 PM Chris Angelico <rosuav at gmail.com> wrote:
>>
>> On Thu, Jun 17, 2021 at 7:35 AM Dan Stromberg <drsalists at gmail.com> wrote:
>> > > >> How well can you define the things you're looking for?
>> > > >>
>> > > >> https://xkcd.com/1425/
>> > > >>
>>
>> > > He means that image processing is a hard problem that requires expertise
>> > > to solve.
>> > >
>> > > >
>> > > > Here is the image, I need to separate the road and markings from the
>> > > rest and divide the image into squares of 100x100 pixels, for each square I
>> > > need to check if it contains a road and markings:
>> > >
>> > > Can you define road in terms of an algorithm that looks at the pixels?
>> > >
>> >
>> > I think that XKCD may be a little out of date.
>>
>> It's not out of date. The task still requires a lot of effort - it's
>> just that the effort is now "preparing a suitable corpus" rather than
>> "figuring out how on earth to do this". Even with all the tools at our
>> disposal, there's still a stark (and often surprising) distinction
>> between the easy and the hard.
>
>
> Well...  Are you sure?
>
> It's no longer a problem that requires 5 years and a research team.

That's because the research years have been done for that particular case.

> It's now a problem that requires hunting for relevant labeled data, and failing that, paying a small team of unskilled laborers minimum wage to classify images from google images or similar.  Plus some programming to create the model and to use it in production.
>
> Deep Learning is catching on, in significant part, because lots of useful data is becoming available.  Also because hardware is getting faster and the algorithms have improved.
>
> Clearly some things are harder than others, but the "hard" example given in the XKCD is no longer really a good example.
>

I don't think it's a bad example. We needed hindsight to be able to
figure that out - see the hover text. And it's still much MUCH harder
to say "is this thing in this person's hand a gun?" than to say "is
this light on?".

But as we've seen from all the various voice-command assistants, the
difficulty of doing a good job doesn't stop people from doing a bad
job...

"Close Spotify." // "Okay. Opening Spotify."

ChrisA