Bamboozled by setup

[bmorgan99]

I've made a deer model in hopes of harvesting veggies before the deer this year, but apparently can't get the config right. It works when I test it via python using the custom url, but in practice deer are rarely if ever are listed in live or testing prediction detail and are "unwanted". These are my AI Server URLS(s):

http://127.0.0.1:8030/v1/vision/detection http://127.0.0.1:8030/v1/vision/custom/deer http://127.0.0.1:8030/v1/vision/face/recognize

Do I link them, set custom as a refinement server, etc.

[VorlonCD]

/detection is the main URL /deer would be set as a "Linked" server (check the link checkbox, uncheck others).
/face/recognize would be set as a "Refinement" server - and make sure PERSON is in the "Use this server ONLY..." list. That way, it wont try to detect the face until a person had been detected in the first place.

[bmorgan99]

Link/Combine and select detection from the drop down or Use Only as linked

[VorlonCD]

Right, to clarify:

• Edit the /Detection URL, click "LINK" checkbox, then put a checkbox next to /Deer.
• Edit the /face/recognize URL, click "Use as Refinement server" and put PERSON into the list below it.

[bmorgan99]

When I run a deer test image from the /detection settings, I get a long delay, no deer hits and a "linked AI server timeout".

[VorlonCD]

the custom models will take quite a while to respond the first time or two as I understand. When you edit the AI server, increase the "maximum time to wait" from 10000 to 30000.

Also try disabling everything BUT the deer custom model, and make sure none of it's checkboxes are enabled before your next test. You have to make sure that is working correctly first.

[bmorgan99]

I restarted ds and ait and went to bed. This morning I have deer hits! Not actual deer, but the same stuff ds would catch and mislabel; spiders, raccoons, cats, etc. I don't see any of those ds guesses in prediction details only one "deer" in the 3x% range, so that is a little worrisome. Could have been the long first attempt thing or just the need for reboot magic.

Thanks.

a sample of their persistence

[clarkhaskins]

Take the pictures that have deer that were not identified properly and add them to your model and retrain. It will take many iterations to get it right.

How many photos did you use to train?

On Tue, Apr 20, 2021 at 6:35 AM bmorgan99 @.***> wrote:

I restarted ds and ait and went to bed. This morning I have deer hits! Not actual deer, but the same stuff ds would catch and mislabel; spiders, raccoons, cats, etc. I don't see any of those ds guesses in prediction details only one "deer" in the 3x% range, so that is a little worrisome. Could have been the long first attempt thing or just the need for reboot magic.

Thanks.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/VorlonCD/bi-aidetection/issues/234#issuecomment-823237099, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADA5A2FJWRLYSK2MM4VOM23TJVYI3ANCNFSM43GTPL4Q .

-- --

Clark Elliott Haskins III

Mobile: 505.385.1484

@.***

https://www.linkedin.com/in/clarkhaskins

There is no place like 127.0.0.1

[bmorgan99]

I have @600 images and I'm on the 9th version doing what you described and also flipping the images I had horizontally and adding. All the mislabelling from last night turned out to be an iris, which does kinda look like antlers at night. My main concern now is that nothing was listed from the built-in ds models. I watched 2 squirrels get completely ignored which are usually caught. As I typed this 2 birds were snagged, so it is working a bit.

I believe that is a Luna Moth top left and that is usually tagged as a bird, but not this time.

[bmorgan99]

Config works. I have unrelated issues.