[Dprglist] Would it be possible to make an object detector that works on any object?

Wed Oct 6 15:16:48 PDT 2021

In some ways yes and in some ways no. At this very moment I'm pursuing the
"solve" through classic vision processing algorithms. Meaning
explicitly constructing all the steps in a processing pipeline with me
(arguably a human) deciding which techniques to use. And the method I'm
focusing the most on is explicit "background removal" based on having a
reference image of an empty room. So I'm removing the floor to isolate
specific obstacles.  It's not an absolutely necessary step to do this since
there are a myriad of other ways to identify connected pixels that
represent a surface that might be in the way. But it is an easy approach
and is facilitated by the geometry of my camera setup.

When I showed this a couple of weeks ago, I kind of raced through it as
more of a status update. I'd be happy to share a more thorough and
explanatory walk through at a future monthly. But this Saturday I'm
committed to a robot team meeting.

I'm tempted to call the OAK-D-Lite update kismet or serendipity or
coincidence, but it's none of those. Vision tech and deep learning are just
maturing and converging at a rate where it's becoming affordable and
accessible to a rapidly widening circle of people.  So more are asking
similar questions.

The note in the kickstarter is different from the approach I'm taking,
maybe? The trick with deep learning is that you're not really sure.  Let's
call this Generic Object Detection. The examples shared are using a
published Tensorflow-lite  model called "Mobile object_localizer_v1.1"  But
there is no description of how it was trained. Most of the objects
identified appear to be nicely segregated from their surrounds and the
surrounds seem to be fairly uniform. So, my guess is that it's effectively
doing background subtraction in a computationally different manner. Could
also describe it as "some of these things are not like the rest"

The trick is, is it doing it better, and is it as useful?  I suspect it's
not doing what I want. Object detection is not what I'm after right now.
Obstacle detection is different enough to be careful of the terms. Think of
obstacles being things that I can run into, but they may only be a part of
a much larger object. Think of being near a large chain link fence or bushy
tree. Not something I'll run into immediately, but maybe they fill the
whole frame. Will a discrete depth-blind object detector work? I won't know
until I try it and I'll do so if my current approach proves insufficient.

I wish folks who publish sensing algorithms would also routinely document
the common failures of the technique. No systems are perfect. Including
humans. We wouldn't have performance magic if our senses and interpretation
were infallible.

Best,

Karim

On Wed, Oct 6, 2021 at 12:33 PM David P. Anderson via DPRGlist <
dprglist at lists.dprg.org> wrote:

> This very cool.   Maybe what Karim was suggesting last evening?
>
> thanks
>
> dpa
>
>
> On 10/6/21 8:35 AM, Doug Paradis via DPRGlist wrote:
>
> * [EXTERNAL SENDER]*
> This is an update from the OAK-D-Lite Kickstarter. It is a really powerful
> technique to help retrain models. It may also answer some of the issues
> brought up in last night's RBNV.
> Regards,
> Doug P.
> Would it be possible to make an *object detector* that works on *any
> object?*
>
> We've been curious about this for quite a while.  As it would be so useful
> (more on that below).  And now thanks to the Open Source community, *you*
> *can* (here
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2FBw_yO9xCSz00YRGOVDw7YcH4h5tO0O1SAC0NVBAfvA1Z_S_JLlQ5b46aIXaAr093vIzWv-4M2WJarxQhBa8R_QexymIi03E4FMkUG1dbG6bv7djvOuDfkKqJ7Kw7Wfq6zm3VVk1w-xbQuMg7rjcCKi9Tb_kscya9reEW2m35SvjGoC5ihsYyt__3jy5yMSlFfM8HPLZZPzF_RWpSayFJlzox6SHDgmgcVthBpYGyrD4NrTUbmnVxOUOykqNn3mIv%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh4%2FZPSlZoHjmpA_tLu2-CTgjEsJLKCgmExBwnmStlLX52M&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107908178%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=JoE%2FINYefhib30fMAUGRlEziAYBSwKeYgLUp41ohimg%3D&reserved=0>)!
> And here it is *running on OAK-D-Lite*:
> Detect Objects Never Before Seen
>
> Normally, you train on *specific classes* - e.g. *banana*, *apple*,
> *raccoon*, etc.  But what if you don't care *what* an object is - just
> that *it's an object*?
>
> We'd been asked this question a lot, *but we didn't have a solution*.
> Heads-down on OAK, we thought to ourselves that this is *probably* possible
> and really wanted to see it happen, but hadn't pursued it.
>
> And then *boom*, *PINTO0309*
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2FBw_yO9xCSz00YRGOVDw7YTtz_W7d9lMX1zeJEFSoBF6iIp3MtKHlCO8bMcMDYlHpBum4rBCgS-kC-lxarWJQZbo9M_BUHO9iD1bmdBL6_R4vDROHqO6N_mdHZNOADSaGO0Fvnyy7FRYVk9PncFb5rQ%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh5%2FuJGbjOsxNegKyTm1lMv0oPc2wafM48qa-tFtsDxcMdc&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107918168%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=B2OO4EQOCVQdxSpXOgoUt4dj7xNvny3G%2B%2FpW8yyVVtU%3D&reserved=0>
>  and *ibaiGorordo*
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2FBw_yO9xCSz00YRGOVDw7YUIYDopaxhYvdgFINFQoskvTkx5CITQOJjvZv8JTn-aYlgTI3NNOxETrS1DY1rtqXkFgTkln-8ErZigZLq8I3Jl6fMDeITSNQ8aCsu_FJqLWw6lQELRPGkAyocaINjhElh4tbG_NXiKjsQaDATDsIGY%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh6%2FUEWCkQAZO_sNKTUc7OQkpnFMLZI_-Pu-rQTcrT1ygaM&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107918168%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=fGT85qmsCrb041EPLn1nU8bvJzrpTlzPoodsmJgcQls%3D&reserved=0> shared
> this in our Discord (*here*
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2F_7Q438LFUsj-9qXXW-Uam_3-0AzMZMqMogxcrPY8ZJfLgHVyFW6J5seQjaTpnM4V%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh7%2FX3H2HJJcG10xM8Pgdgrj1gb8mPW7Wj1z0FU64DKErt0&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107928168%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=ljhGG%2F9HRq07Ro4o8%2Fdkam%2Fnnubfk%2FGAos1%2FA6QcujI%3D&reserved=0>)
> with the solution below 5 days after our campaign went live.
>
> And if you check out ibaiGorodo's Github dedicated to this (here
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2FBw_yO9xCSz00YRGOVDw7YUIYDopaxhYvdgFINFQoskvTkx5CITQOJjvZv8JTn-aYlgTI3NNOxETrS1DY1rtqXkFgTkln-8ErZigZLq8I3Jl6fMDeITSNQ8aCsu_FJqLWw6lQELRPGkAyocaINjhElh4tbG_NXiKjsQaDATDsIGY%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh8%2F-pD5JhK7CEhn1lkPSlXUG66r4BEKLM-WNokrvPnYUhM&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107928168%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=Uuct%2Fsz6rRQG1yigdnMjBHyN9WMeY%2FkIG0dJ2UYQ10k%3D&reserved=0>
> ) *you'll notice something familiar*:
>
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2FBw_yO9xCSz00YRGOVDw7YUIYDopaxhYvdgFINFQoskvTkx5CITQOJjvZv8JTn-aYlgTI3NNOxETrS1DY1rtqXuufPFjhcn9z1w7_89OvCMA%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh9%2FNM-jpKlFYjUIpKzKblikHleHa031y81Lwvyh2STC-EY&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107938153%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=F2yXkSgh3BMULsld7Z1nPXaNHzAq3QI8rZhjfqr95hQ%3D&reserved=0>
>
> He actually used our campaign image for *OAK-D-Lite as his test sample*!
>  (And gave a shout-out to the campaign, which is much appreciated!)
>
> And this class-agnostic object detector *correctly detects OAK-D-Lite as
> an object*.  Even though this model had for-sure had *never seen
> OAK-D-Lite before*.  So, it's working!
> And It Works on All Sorts of Things
>
> Thanks to PINTO0309
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2FBw_yO9xCSz00YRGOVDw7YTtz_W7d9lMX1zeJEFSoBF4u2trM4ihnQS2Zjkz7L8Mu%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh10%2FgMoHDxTNPjI2eBa3QcVpwRu2uj_0K6BovaXnFq7DeIA&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107938153%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=5bQqwJRZRI1oJ%2FiT61MUHwOQ%2FE33mFCSQljQK3wHIs8%3D&reserved=0>'s
> conversion of the model to run on OAK-D-Lite, you can try it out when your
> unit arrives.  And in the meantime, you can check out the crazy number of
> models that PINTO0309 has found, and done the hard work of converting, in
> his model zoo, here
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2FBw_yO9xCSz00YRGOVDw7YTtz_W7d9lMX1zeJEFSoBF6iIp3MtKHlCO8bMcMDYlHpJyuxEEu9pSSzbkfqAkRcJw%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh11%2FI9tT1fK3qacYVgrp0sBLL_qipQ9ZR3CSZdO-s5GqEWA&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107948147%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=eyoDHyLPJ7xcw4TSdrv86gTyni4bCkyy4cCleUT3sVE%3D&reserved=0>
> .
>
> And then also thanks to IbaiGorordo
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2FBw_yO9xCSz00YRGOVDw7YUIYDopaxhYvdgFINFQosksgPR3EGAUw09sqtQNbZE0D%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh12%2FqKtn3GcjJqPr_VWLaHadDFshkcVmddQXYSB7y50Onoc&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107948147%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=RH3a5oXGN%2FL54hzWsVXBGfnsQAT%2Bst9mE5bMbP%2Bm%2F1I%3D&reserved=0> for
> sharing his experiments with this model, here
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2FBw_yO9xCSz00YRGOVDw7YUIYDopaxhYvdgFINFQoskvTkx5CITQOJjvZv8JTn-aYlgTI3NNOxETrS1DY1rtqXrcKvgbI4QOrNF24krtLFsJAwAvs30sxZohsbq8Z1zll%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh13%2FPtDPBB4nOFqiiuomSnHI93BVZsqyOBn_FQouimiTO98&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107958146%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=oyI%2Fmj%2FG5l2idcZ6BLl49H81ZHiMKgQ%2FcheVm0ZQpgE%3D&reserved=0>,
> which we'll reproduce (plagiarize) below - as they're super cool examples
> that show how well this model works.
>
> *A Cool Stop Motion Series*
>  Original video by Animist: https://youtu.be/uKyoV0uG9rQ
>
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2F5Ixklp3QcabAdTzhlCDOgiJ3lFAafZF0Zirf8MCKZfjZlsmjQ58WjfMErmW0H_RD%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh14%2FXArQX7odP-apBtdXaIeeYiZ2S9eBk8dNpM4yhP_lGHA&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107968145%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=UBJTlJbcxNW6I3CpCCUi6SLvKFAUdRArYPFNCT1IXto%3D&reserved=0>
>
> *A Cabybara*
>
> He could have made this up and I (Brandon) would have no idea if he did.
> So I at least am no smarter than this class-agnostic object detector in
> this case.  I know it's a thing, but not what it is.
>  Original image:
> https://commons.wikimedia.org/wiki/File:Capybara_portrait.jpg
>
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2F72XxU03wGIudEQFoB9UCGrCv2c4mwxkrrV_78a2oJwOh2OGXaaX4Z5UeMe80Vmj_i-lMFLVAveYsmA32K5dFpsA7Z86OX631ri8KrxF0Rjc%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh15%2FpTWMPwk6uw8jAhBnq_Zo1mLsIbRQgzHHkDWkdt1iV1w&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107968145%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=7Mkn7Tqn5bwqID34qEF7%2FsSx60%2FVy%2BKUT3gzlXGgxco%3D&reserved=0>
>
> *Coins*
>
> All of which I've never seen.
>  Original image: https://commons.wikimedia.org/wiki/File:Japanese_Coins.jpg
>
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2F72XxU03wGIudEQFoB9UCGrCv2c4mwxkrrV_78a2oJwOkfCuRxPaHhraq2I294XtNrEKTQI-tq6X52rkHD6pRtSUW_FCblH10oetgz53jhgI%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh16%2FZJBrmEuaC38PPSzdXp11OSBMxSKW82kq_e0x5dTMurY&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107978131%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=hX%2FJb8aNLoioLbhE6y0isi8XHkImU0P25oiu2zGafPY%3D&reserved=0>
>
> *Shoes*
>  Original image: https://commons.wikimedia.org/wiki/File:Japanese_Coins.jpg
>
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2F72XxU03wGIudEQFoB9UCGrCv2c4mwxkrrV_78a2oJwOkfCuRxPaHhraq2I294XtNrEKTQI-tq6X52rkHD6pRtSUW_FCblH10oetgz53jhgI%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh17%2F6ZO0bsiYfJ0fCKNPb9wnITMSq5aoSFrEVh3X2hiKT_E&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107978131%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=30QsF5TW8G2ZGKXHjU2c004y8zIOIn%2B1%2FbgAy1pwvpg%3D&reserved=0>
>
> *Some Cool New SpaceX Thing*
>
> Is that a rendering or real life?
>  Original image:
> https://en.wikipedia.org/wiki/Spacecraft#/media/File:SpaceX_Crew_Dragon_(More_cropped).jpg
>
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2F-QYxrLLAE-TM8D3U_fdmudCBIsPA3mViJ_Vf_f20QYOnVcQsEg-gJz2RtsesBoOSDfNUrulnoWQxPMTVMeiKDeW_orL8kfP3J6Wr4lD7ARGRHjSPtIKGJV0lf7UD3lbU_qCTNvAch2222RdS_rMVVw%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh18%2Fbl86J7BKk-fcuETk3XAddkapEQbd7oHNmB0SqS-Y6JE&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107988123%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=X3oLeakfo0%2FNE9WO%2FTVosdzO%2FqhsBBw5vsW9j8fqM8E%3D&reserved=0>
>
> *A Window*
>  Original image:
> https://commons.wikimedia.org/wiki/File:Window_-_Paddington_-_London.JPG
>
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2F72XxU03wGIudEQFoB9UCGrCv2c4mwxkrrV_78a2oJwOtonLmjpc-h29l0W7wvRQmhveEyk_QmrRRx70ijB8MbN_3JKIfUY_u1AnDGqktg9A%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh19%2Ff3gqUvNacYn10ZuCtY4TMyZ09EjdTkVYQEgI2RksiSc&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107988123%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=5bqP%2B%2Bv8VivfKkKBhfgUcVa6Ywojqxh%2Be0u1JAkX1oc%3D&reserved=0> And
> Now To Play with it on OAK-D-Lite!
>
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2FpGVEG5T9w7LP_74vNtZiwKfPhhXVwL4vO_Xv8CXD7oT9tNJ1cyb_SB-ESaukqjlk2dC9t3GN2qD6c7yOJJw1Sodu1JzOiWPlROM9LJ8t2JF5Z45NV4HuOsRpU51I7l3yjncDpaqsCgQrllCGmIqpXMUECT_V3DOFxz-g06abXEF5uMcvaB9R5CSoVL0A764M3GSPML7O3Xd2CXFVQv7brlZcmKgVphUjd-56VdQlSk6Yho7phqXF3V7XFFL4TQ2q%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh20%2FFPPjN2TCgUmQcUehR_c9OGFjMLdRwkll4AFxhoyzXNI&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107998118%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=4ArdBwWU69ZvaII3JBgxKYUe%2FT3wEgNj40w%2FsODB%2Bxw%3D&reserved=0>
>
> You can see it's not perfect.  But it sure does pick up a lot of things.
> It even picked up the glasses on my face when trying it out:
> And Why Is This Useful?  Model Re-Training.
>
> This solves one of the classic *chick-and-egg problems* of deploying
> custom-trained models in the wild:
>
>    - What if my model *isn't picking up objects that it should*?  How do
>    I know this?
>
> This is a painful problem.  Particularly if you are deploying on the edge
> - where you probably don't want to be sending back tons of data to the
> cloud to check.
>
> Having a generic object detector is a *hugely beneficial piece* to
> solving it!
>
> Here's the idea:
>
>    - *Run this generic model in parallel with your custom model* (recall
>    on OAK-D-Lite you can do arbitrary series/parallel combinations of neural
>    models - only limited by DDR bandwidth/size).
>    - *Monitor* on-device* when the generic model detects things your
>    custom model doesn't*
>    - *If/when* this happens, *save/send the cropped results from the
>    generic detector* back for analysis and potentially re-training your
>    custom model.
>
> And best of all, since OAK is plug&play with *Roboflow*
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2F_7Q438LFUsj-9qXXW-Uam2QmTB0cC_vLPCsUDYzwDgUUW4tiSJ9dR0XuHH1cuGVCnz5yV7TzsFU1zc57UzbFPA%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh21%2FxuQXmr1AZvariQpIDuyKduYq1a4aFegBzuAx8NBqiX0&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942107998118%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=KqDQgyFzOPgsEpJUCLYbT0wXAZt05f5lNG9sc2%2FDSMw%3D&reserved=0> -
> you can automate all of this with their *Upload API*
> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Femails.kickstarter.com%2Fss%2Fc%2F_7Q438LFUsj-9qXXW-Uam2QmTB0cC_vLPCsUDYzwDgWKPqhzZpsmw297CfnKNeJmo4Lnm0yo24SiSe2q0w138Q%2F3fx%2FYEU8u-ZJQ9iOQlUeuhziZA%2Fh22%2FXyG7KwGgWD9M4HOZPEq4Tw1sChFghvO96-3b7mYFY-Y&data=04%7C01%7C%7C1c6128d537ad48e2e40f08d988887efb%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637690942108008119%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=eS7UY6xrJqohy0P6suCJ28siEbbic3fxCfevTZqY3zo%3D&reserved=0>
> .
>
> Pretty cool.
>
> *Get feedback automatically on all the objects your model **isn't*
> * detecting* - have them *uploaded automatically* so you can label the
> ones that should be detected - and *retrain your model over time to
> continuously get smarter from this*.
>
> That's all for today.  Likely some other cool thing will be done with OAK
> by tomorrow that we can share!
>
>  - OpenCV/Luxonis Team
>
> _______________________________________________
> DPRGlist mailing listDPRGlist at lists.dprg.orghttp://lists.dprg.org/listinfo.cgi/dprglist-dprg.org
>
> _______________________________________________
> DPRGlist mailing list
> DPRGlist at lists.dprg.org
> http://lists.dprg.org/listinfo.cgi/dprglist-dprg.org
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.dprg.org/pipermail/dprglist-dprg.org/attachments/20211006/792c89d1/attachment-0001.html>