5.5.8 - June 13, 2022 - Code Project’s SenseAI Version 1 - See V2 here https://ipcamtalk.com/threads/codeproject-ai-version-2-0.68030/

[clk8]

Going to have a super long queue if you process a lot of images per trigger. I do 20 images on the doorbell and 10 on the street. Sometimes cameras trigger at the same time within quick succession.

This is interesting...why do you process so many images?

I can see this leading into a very long conversation on these settings. I believe i have a very non-traditional set up for this. I turned my max alert duration all the way to 10sec. I think I am only processing 5 images on most of my cameras and 10 on some...not even sure why other than I watched a video when I first set it up that suggested 5

 

[sebastiantombs]

Every second is power consumed. GPUs are not as efficient, power wise. If you're happy with that and the potentially long queue times that result than it's good.

 

[clk8]

Yeah, it's not the cleanest way to disable the ai. It should get better once BI gets better integration with Sense ai.

I am pretty sure this means it is only using yolov5x

NOTE: It identified my van as a train

 

[tofu]

This is interesting...why do you process so many images?

I can see this leading into a very long conversation on these settings. I believe i have a very non-traditional set up for this. I turned my max alert duration all the way to 10sec. I think I am only processing 5 images on most of my cameras and 10 on some...not even sure why other than I watched a video when I first set it up that suggested 5

I process more images for better accuracy. Doorbell is most important to me because it handles the facial recognition. Even with 20 images it still misses sometimes!

 

[TBurt]

View attachment 137709

I am pretty sure this means it is only using yolov5x

NOTE: It identified my van as a train

I will take a van as a train over a stray dog sending me alerts while I sleep saying there is a person at my front door at 2am. That actually happened, Or better yet a tree as a person. I have not had any tree people (even without masking off the areas) sense moving to 5x. I am sure we will find a tree person eventually but at least not all trees are people. It's not perfect but a big step in the right direction. It seems to think cats are dogs, like 80+ % are detected as dogs. At least my dog is no longer a cat, or worse a bird now. I have noticed yolov5l seems to get dogs and cats more accurately than 5x so far in my testing. Now we need super Mike to work his magic and blend general with Yolov5x and Yolov5l. That will be another big step...

 

[clk8]

I process more images for better accuracy. Doorbell is most important to me because it handles the facial recognition. Even with 20 images it still misses sometimes!

What doorbell are you using? I currently have a Ring so it is not in my BI

 

[bhedrington]

View attachment 137665

The top post should have it all listed:
(1) IP Cam Talk Custom Community DeepStack Model | IP Cam Talk

More tech info:
GitHub - MikeLud/DeepStack-Security-Camera-Models

Thanks,

I did see that post, but ignored it as im on CodeProject AI not on deepsrtack

Perhaps whoever posted that can revise the post and put a not there that says it is for codeproject as well

Perhaps a text file on the github with the list of the labels for the downloads?

I'm sure over time the labels included will cahange.

 

[TBurt]

Thanks,

I did see that post, but ignored it as im on CodeProject AI not on deepsrtack

Perhaps whoever posted that can revise the post and put a not there that says it is for codeproject as well

Perhaps a text file on the github with the list of the labels for the downloads?

I'm sure over time the labels included will cahange.

Yeah, I can see home that would be really confusing to someone that hasn't followed both threads, and CodeProject + Deepstack.

I agree the poster should revise the post. MikeLud, please see to it ASAP! No, wait do it right after you start the Yolov5x training of your general model. LOL I am kidding Mike. I know home busy you have been with work and helping us.

 

[MikeLud1]

Yeah, I can see home that would be really confusing to someone that hasn't followed both threads, and CodeProject + Deepstack.

I agree the poster should revise the post. MikeLud, please see to it ASAP! No, wait do it right after you start the Yolov5x training of your general model. LOL I am kidding Mike. I know home busy you have been with work and helping us.

I will update the post tomorrow

 

[CrazyAsYou]

I keep wondering... why do I care if it takes 100ms or 1sec for the AI to process an alert? Isn't accuracy more important and how does the system taking a second or two to register an alert affect my security? Am I missing something?

As others have said already, not only can it lead to a long queue of images waiting to be processed on busy systems and extra CPU/GPU power consuming cycles but also many of us (me included) have a setup where triggers lead on to some form of home automation for things like a doorbell system. I feed into Home Assistant using mqtt to then turn lights on, performs Alexa announcements and Pushover notifications, I like to get this feed very quickly while someone is on approach to the door and not 5-10seconds later as they already leaving – I also want to process ~20 images quickly (at least 10 per second) to reduce false alerts.

 

[TBurt]

Mike,
Could you shed some light on what these different settings for sense ai do? I get the resolution, but what are the resolution of images are fed to the AI models for low,med,high? I tried high with yolov5x and I was getting person detections of 70 to 80% of people inside their cars, behind the glare of their wind shield. It was awe inspiring, but my detection times dropped to 400 to 600ms. So, I went back to med. If I only had 2 or 3 cams, that would be my go-to choose.
Model size I am not sure I completely understand what those changes do, so I left it alone.

"MODEL_SIZE": "Medium", / small, medium, large, x-large
"RESOLUTION": "medium", / low, medium, high

I think I found part of my question out. Resolution = Mode in Blue Iris, and Deep Stack. So, project code is running on medium settings compared to high DS was using. This would explain how project code is so fast. Or part of it anyways. I think sense ai is faster regardless of the mode BI is sending to the AI. The settings above is for custom models. This screen shot is for built in model.

 

[slidermike]

I am unsure how to tell BI to utilize more than 1 custom model set in the AI section.
Do we have to put a ; or some other delimiter to tell BI that its more than 1 custom set?

And where are the YOLO5 .pt files?
Sorry for the newbie questions but finding this after the fact is not as straight forward as if you are here the whole time and following the threads as they are made.

Thank you

 

[MikeLud1]

Below are the labels for each of the models

CodeProject.AI-Custom-IPcam-Models
IPcam-combined Labels: - person, bicycle, car, motorcycle, bus, truck, bird, cat, dog, horse, sheep, cow, bear, deer, rabbit, raccoon, fox, skunk, squirrel, pig

IPcam-general Labels (includes dark models images): - person, vehicle

IPcam-animal Labels: - bird, cat, dog, horse, sheep, cow, bear, deer, rabbit, raccoon, fox, skunk, squirrel, pig

IPcam-dark Labels: - Bicycle, Bus, Car, Cat, Dog, Motorcycle, Person

 

[MikeLud1]

I am unsure how to tell BI to utilize more than 1 custom model set in the AI section.
Do we have to put a ; or some other delimiter to tell BI that its more than 1 custom set?

And where are the YOLO5 .pt files?
Sorry for the newbie questions but finding this after the fact is not as straight forward as if you are here the whole time and following the threads as they are made.

Thank you

To use more the one model add model like below in the camera AI seettings

 

[CrazyAsYou]

Has anyone with a 4GB RAM GPU got SenseAI working (The GPU version) ?

 

[TBurt]

I changed the settings to high, and x-large to better match was DS and BI used. I was using yolov5l model also. I let it run for some time. Just watching it I could tell it slowed down some, but I was not sure, so I left it to do its thing. I checked the logs later and it went from low 60ms to 70 to 75 max on average for my system. I doubt Deep Stack would do Anyware close to that using these models. DS would have most likely started giving me AI is not responding, and then lockup for a while or just crash outright. I was running yolov5s on DS, and it was slower. s is faster but less accurate in my limited testing compared to L and x. I am even more impressed now. Now they need to get it more user friendly with the integration to BI. Right now, I am having to manually edit everything little change. Most people will not or won't be able to do this. I am happy to also report that using high and x-lage that the L model and x both did not miss any detections. That they matched ipcam general on fast moving cars, people, and what not at night. General was the only model that I could run by itself and be confident that it was not missing detections. I am not sure how much of that is from the models, and home much was from the new Ai though. I am happy to see the new ai coming along and working this well in regard to detection accuracy/misses/etc being so new. I am also not getting the random Ai not responding errors/Ai crashed using the new AI. It can only get better from here!

I wish I had a second cuda gpu so I could help Mike out with trainings his models using Yolov5x, Yolov5s, and Yolov5l. I may have to make a new purchase soon now that gpus are dropping in price.

"Y ,pde;s/ OLOv5_VERBOSE": "false",

"MODEL_SIZE": "x-large", / small, medium, large, x-large
"RESOLUTION": "high", / low, medium, high
"USE_CUDA": "True",

"APPDIR": "%MODULES_PATH%\\DS-Models",
"MODELS_DIR": "%MODULES_PATH%\\CustomDetection\\assets"
},

 

[Village Guy]

Has anyone managed to run the GPU install script install_CUDnn.bat ?
I tried the latest version today again and failed miserably. Does it need to run in a special environment?
I have tried running in a Windows Terminal, Powershell, CMD and direct execute with Admin privileges.

 

[MikeLud1]

Has anyone managed to run the GPU install script install_CUDnn.bat ?
I tried the latest version today again and failed miserably. Does it need to run in a special environment?
I have tried running in a Windows Terminal, Powershell, CMD and direct execute with Admin privileges.

The script worked for me,
Does the PC you are running the script have internet access? It needs internet access to download the required files.

 

[misterfredsr]

I've been running IA for a few days now and WOW what a difference it makes. I have a PTZ watching my back yard and every other minute I would have a bug trigger my cam and now with IA I don't get any alerts at all from bugs. Going to have to make some time to play with this some more. I have an acre aspire TC-710, i7 with a @ 3.4ghz and my CPU is running around 20% with 9 cams and when reviewing with IA testing on the CPU jumps up between 80% and 90%.

 

[wittaj]

I've been running IA for a few days now and WOW what a difference it makes. I have a PTZ watching my back yard and every other minute I would have a bug trigger my cam and now with IA I don't get any alerts at all from bugs. Going to have to make some time to play with this some more. I have an acre aspire TC-710, i7 with a @ 3.4ghz and my CPU is running around 20% with 9 cams and when reviewing with IA testing on the CPU jumps up between 80% and 90%.

Are you running substreams? 20% seems high for what I believe is a 6th gen CPU?