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/

[CrazyAsYou]

Add a way to disable modules from the dashboard - I am sure they will implement this in the up come week or so.

Better yet a custom options section during install so we can unselect the modules we don't want and also the custom models we do/don't want. I'm perfectly happy editing json files and deleting models but many people might not be.

 

[Vettester]

Better yet

And what happens after the install if you change your mind and want to add or remove modules and/or models? The web UI option gets my vote.

 

[CrazyAsYou]

And what happens after the install if you change your mind and want to add or remove modules and/or models? The web UI option gets my vote.

same as many other installers, you run a change reinstall option, why download and use disk space for things that might never be used?

 

[Vettester]

same as many other installers, you run a change reinstall option, why download and use disk space for things that might never be used?

On my CPU version a reinstall takes about 20 minutes and the file structure for the custom models is 281MB. The space used for the additional models is minimal in comparison to the value of time. If you're really concerned about disk space you can always go into the asset directory and manually delete the models you know you'll never want to use.

 

[MikeLud1]

CodeProject.AI and Blue Iris integration update

 

[toastie]

I don't know what to make of my findings.
DeepStack was previously uninstalled, though a search show there are a few remnants of DeepStack as in Users>MyBlueIris>AppData
I have been on Code Project AI 1.5.6.2, and playing an Alert shows no squares formed around objects in Testing & Tuning whether or not I have in Global Settings AI "Default object detection" selected.
When I was using Deepstack until a BI update messed things up perhaps, squares and oblong were drawn around objects detected by Deepstack's AI. It was working, it deleted traffic headlights Alerts which would have otherwise triggered recordings. It was rather good at this.

This afternoon I uninstalled 1.5.6.2 and installed 1.5.6. I have got back squares and oblongs being drawn around objects but only if I select Default object detection. It showing up many potted plants and a bowl which suggests that it's using the DeepStack data set.

Looking in Blue Iris Status in the AI Tab, it seems to be sort of working. I only have two entries on the left, a red icons for the camera and one for a car it has detected. On the right in the log it's good at finding potted plants, though I do notice that it is looking for a license plate which I didn't ask it to.

It seems here that with Default object detection Code Project AI 1.5.6 is using both the DeepStack data set and the Custom data set irrespective of my wishes. I only asked for Custom Model ipcam-general in the camera settings that I am using for the camera being used here.

Deselect Default object detection and no squares or oblong are being drawn as they were previously. Should I expect Code Project AI to draw squares and oblong around objects it detects in the same way DeepStack does?

In the Status log file, on the left two red icons, one for the camera and a red X, no car as the icon was previously when Default Object Detection was selected (and I also had squares and oblongs being drawn). It looks like my DeepStack data set remnants are finding a car whilst the Custom data set associated with Code Project AI isn't.

On the right it looks like AI has used these data sets,
"api"."actionnetv2".
"api"."ipcam-animal".
"api"."ip-cam-combined".
"api"."ipcam-general.
"api"."license-plate".

It reports "found", "access true","predictions", but everything else is blank.

This is beyond my understanding.

 

[MikeLud1]

I just submit a feature request to add a way to disable modules from the dashboard.

View attachment 137426

Update on this feature request

 

[IAmATeaf]

I don't know what to make of my findings.
DeepStack was previously uninstalled, though a search show there are a few remnants of DeepStack as in Users>MyBlueIris>AppData
I have been on Code Project AI 1.5.6.2, and playing an Alert shows no squares formed around objects in Testing & Tuning whether or not I have in Global Settings AI "Default object detection" selected.
When I was using Deepstack until a BI update messed things up perhaps, squares and oblong were drawn around objects detected by Deepstack's AI. It was working, it deleted traffic headlights Alerts which would have otherwise triggered recordings. It was rather good at this.

This afternoon I uninstalled 1.5.6.2 and installed 1.5.6. I have got back squares and oblongs being drawn around objects but only if I select Default object detection. It showing up many potted plants and a bowl which suggests that it's using the DeepStack data set.

Looking in Blue Iris Status in the AI Tab, it seems to be sort of working. I only have two entries on the left, a red icons for the camera and one for a car it has detected. On the right in the log it's good at finding potted plants, though I do notice that it is looking for a license plate which I didn't ask it to.

It seems here that with Default object detection Code Project AI 1.5.6 is using both the DeepStack data set and the Custom data set irrespective of my wishes. I only asked for Custom Model ipcam-general in the camera settings that I am using for the camera being used here.

Deselect Default object detection and no squares or oblong are being drawn as they were previously. Should I expect Code Project AI to draw squares and oblong around objects it detects in the same way DeepStack does?

In the Status log file, on the left two red icons, one for the camera and a red X, no car as the icon was previously when Default Object Detection was selected (and I also had squares and oblongs being drawn). It looks like my DeepStack data set remnants are finding a car whilst the Custom data set associated with Code Project AI isn't.

On the right it looks like AI has used these data sets,
"api"."actionnetv2".
"api"."ipcam-animal".
"api"."ip-cam-combined".
"api"."ipcam-general.
"api"."license-plate".

It reports "found", "access true","predictions", but everything else is blank.

This is beyond my understanding.

Currently I thought that all the custom models are installed and you have remove/move the ones you didn’t have want to use?

 

[tofu]

Personally, I think one of the most exciting features down the pipeline is coral tpu support. That'll be a major power savings and will open up more use cases for micro-pcs that can't fit discreet gpu

 

[IAmATeaf]

Personally, I think one of the most exciting features down the pipeline is coral tpu support. That'll be a major power savings and will open up more use cases for micro-pcs that can't fit discreet gpu

I’m waiting to see if the onboard GPU can be used, with BI not really needing it I’m hoping that they can make use of it to improve detection times, think on my i5-6500 they are around 500ms plus.

 

[TBurt]

Personally, I think one of the most exciting features down the pipeline is coral tpu support. That'll be a major power savings and will open up more use cases for micro-pcs that can't fit discreet gpu

My BI server is a mini-pc, ASRock DeskMini X300. I used one of its m.2 nvme slots to add a full size pci slot. The adapters are not too expensive, and they work well. The gpu a GTX 1060 sits next to the pc on a stand. I can run any Gtx 7 series gpu on up to RTX. 1060 was good for power draw to performance IMHO. Definitely better than buying a different pc.

 

[TBurt]

Just tried the new SenseAI version on my Jetson Nano. Not working and I guess it will never work, as Jetson Nano only supports Cuda 10.2.

Just curious if it's worth to ditch the Nano with a decently running DeepStack detection to switch to SenseAI on a I7 desktop PC... I can't see any obvious improvements... I'm mostly interested in reduce false positives, that with DS and MikeLud1 custom model are frequent ...

I have been playing with the model yolov5x and it is so far the most accurate model. It really hits my gpu. It uses 50+% of it compared to 10% on most. No trees as people, dogs as people, etc so far. There are others less taxing to try if it too much for someone's system. You can get the .PT file here:

GitHub - ultralytics/yolov5: YOLOv5 in PyTorch > ONNX > CoreML > TFLite

YOLOv5 in PyTorch > ONNX > CoreML > TFLite. Contribute to ultralytics/yolov5 development by creating an account on GitHub.

github.com

 

[MikeLud1]

I have been playing with the model yolov5x and it is so far the most accurate model. It really hits my gpu. It uses 50+% of it compared to 10% on most. No trees as people, dogs as people, etc so far. There are others less taxing to try if it too much for someone's system. You can get the .PT file here:

GitHub - ultralytics/yolov5: YOLOv5 in PyTorch > ONNX > CoreML > TFLite

YOLOv5 in PyTorch > ONNX > CoreML > TFLite. Contribute to ultralytics/yolov5 development by creating an account on GitHub.

github.com

What GPU do you have and what model were were you using before trying yolov5x and the detection times?
For yolov5x what detection time are you getting?

I might try retraining my models based on yolov5m yolov5l & yolov5x

 

[TBurt]

What GPU do you have and what model were were you using before trying yolov5x and the detection times?
For yolov5x what detection time are you getting?

I might try retraining my models based on yolov5m yolov5l & yolov5x

I have gtx 1060 3gb. I was using the general model and combined depending on if I would be alerted at 3am or not on a cam. General was good at catching everything but high false positives. Great for cars and streets. Detection times Sub 90 ms. I only use it on my front/back door and the cam on my cars. Speed comes second to accurate detections. I've been looking at the yolov5l also. Nice tradeoff for resource and accuracy/speed. yolov5m doesn't seem to like sense ai in my machine. DS didn't mind it.

 

[TBurt]

What GPU do you have and what model were were you using before trying yolov5x and the detection times?
For yolov5x what detection time are you getting?

I might try retraining my models based on yolov5m yolov5l & yolov5x

Here are some tests using the sense ai test images. Notice the last ones from "menagerie", general. LOL

 

[bhedrington]

Hello all,

Is there a list of all the items that the ipcam-??? will detect, I got many whilst reading through this thread, but where is a complete list.

Thanks
Bret

 

[toastie]

Quote
"Currently I thought that all the custom models are installed and you have remove/move the ones you didn’t have want to use?"

In a camera's setting you can specify the Custom model, presumably this allows customization, for example some cameras aren't going to be looking at license plates. In my earlier post this camera was only set up for model ipcam-general yet the log showed it was looking for license plates. There seems to be some redundancy if Code Project AI is using all models on all cameras despite only one model chosen in an individual camera's setting.

Currently I am reluctant to delete the DeepStack model data set remnants after I uninstalled DeepStack, I have at least got a partially working AI system, it's just that on it's own the Code Project AI data sets appear not to be working when I deselect "Default object detection". I was rather hoping that my observations might help to discover why others like me aren't getting CP AI to work for them.

edit: typo corrected

 

[TBurt]

Hello all,

Is there a list of all the items that the ipcam-??? will detect, I got many whilst reading through this thread, but where is a complete list.

Thanks
Bret

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

 

[TBurt]

What GPU do you have and what model were were you using before trying yolov5x and the detection times?
For yolov5x what detection time are you getting?

I might try retraining my models based on yolov5m yolov5l & yolov5x

I just rechecked I am getting detection times around 60, and my gpu isn't hammered any longer. The only change I did was turn off sense ai built in model. Weird... If you retrained your model with yolov5x and made it sleek I can only imagine the detection times.

 

[tofu]

I just rechecked I am getting detection times around 60, and my gpu isn't hammered any longer. The only change I did was turn off sense ai built in model. Weird... If you retrained your model with yolov5x and made it sleek I can only imagine the detection times.

My ai times are horrible with yolov5l (1500-2000ms), but it's so much more accurate than ip-combined. I'll just continue to use it this way.

To be honest, sensai is overall much slower than deepstack. I was averaging 40ms with the combined models on DS, and on sensai it's about 350ms. CUDA for both. I'm just too lazy to revert back.