IP Cam Talk Custom Community DeepStack Model

[MikeLud1]

I just uploaded General V2.1 model in the first post. I was able to get better results then V2.0. Originally when I started training V2.1 I was going to run the training 300 cycles but the training crashed after 77 cycles and I couldn't restart the training to complete the 300 cycles. During the 300 cycle training I noticed that the model was maxing out on its accuracy so I only did 100 cycles of training for V2.1

IP Cam Talk Custom Community DeepStack Model

I figured out how to download the images that DeepStack used to train the object model. With these images I can start to make the community custom DeepStack model, below are the steps to create the custom model. The first step would be to create a new DeepStack custom model using the same image...

ipcamtalk.com

 

[sebastiantombs]

I'll switch over tonight Mike. Thanks again for all the work you put into this!!

 

[kc8tmv]

I just uploaded General V2.1 model. I was able to get better results then V2.0. Original when I started training V2.1 I was going to run the training 300 cycles but the training crashed after 77 cycles and I couldn't restart the training to complete the 300 cycles. During the 300 cycle training I noticed that the model was maxing out on its accuracy so I only did 100 cycles of training for V2.1

IP Cam Talk Custom Community DeepStack Model

I figured out how to download the images that DeepStack used to train the object model. With these images I can start to make the community custom DeepStack model, below are the steps to create the custom model. The first step would be to create a new DeepStack custom model using the same image...

ipcamtalk.com

Does “General” cover what your combined and your dark did?


Sent from my iPhone using Tapatalk

 

[MikeLud1]

Does “General” cover what your combined and your dark did?


Sent from my iPhone using Tapatalk

General V2.1 was trained with the original images plus the dark images

 

[sebastiantombs]

I switched to V2.1 yesterday evening and it is detecting quite well, too well though t times. I've had some false positives identifying a squirrel as a person this morning. Not a big deal by any means, just a little annoying when the "hockey horn" goes off for someone in the back yard and it's a tree rat with four legs, not two.

 

[MikeLud1]

I switched to V2.1 yesterday evening and it is detecting quite well, too well though t times. I've had some false positives identifying a squirrel as a person this morning. Not a big deal by any means, just a little annoying when the "hockey horn" goes off for someone in the back yard and it's a tree rat with four legs, not two.

Thanks for the update. I also sometimes get false positives on cats. Below is one from last night. I thought adding cat and dog to the model might help the false positives

 

[Sybertiger]

I switched to V2.1 yesterday evening and it is detecting quite well, too well though t times. I've had some false positives identifying a squirrel as a person this morning. Not a big deal by any means, just a little annoying when the "hockey horn" goes off for someone in the back yard and it's a tree rat with four legs, not two.

I know whacha mean. Time to crank up the minimum requirement higher but 80% is pretty high.

 

[MikeLud1]

I switched to V2.1 yesterday evening and it is detecting quite well, too well though t times. I've had some false positives identifying a squirrel as a person this morning. Not a big deal by any means, just a little annoying when the "hockey horn" goes off for someone in the back yard and it's a tree rat with four legs, not two.

I have cats that think they are tree rats
View attachment Clipboard_Cam8.20220531_035916-035926.64.mp4

 

[jaydeel]

Just FYI, I was repeatedly getting Cat instead of Person (well over 50% of the time) with the new Dark model. I switched back to the Combined model.

 

[sebastiantombs]

I think a lot of these mis-identifications may be somewhat unique to each different situation due to contrast problems that DS has versus the human eye/brain.

 

[jaydeel]

I should have all mentioned that the affected cameras are running B/W, not color at night.

 

[sebastiantombs]

I've also been getting mailboxes identified as people again. That was common with the built-in model, objects, but went away with combined. A quick addition of a privacy screen in the camera cured it though. I'm just surprised that they get identified at all because they never move and should show as occupied if recognized. That's in color, too, but not in B&W which is another head scratcher.

 

[MikeLud1]

I should have all mentioned that the affected cameras are running B/W, not color at night.

I also get some false positives in color at night

 

[mangocrysis]

Mike will you be updating the combined next?

 

[MikeLud1]

Mike will you be updating the combined next?

Yes, I do have some more animal images (mostly deer) to add to both the combined and animal model.

 

[aadje93]

So if i understand correctly, i use general V2.1 model i need to put

bicycle,bus,car,cat,dog,motorcycle,person,truck in the "to confirm"

and put objects:0,general in the custom models row and uncheck "use default object detection"?

 

[MikeLud1]

bicycle,bus,car,cat,dog,motorcycle,person,truck in the "to confirm"

Correct

nd put objects:0,general in the custom models row and uncheck "use default object detection"?

If you uncheck "use default object detection" you don't need objects:0 just general

 

[jaydeel]

Alternatively, if you always use "objects:0", you'll never have to discover that "use default object detection" was inadvertently toggled on.
(it happened to me ).

 

[aadje93]

Correct

If you uncheck "use default object detection" you don't need objects:0 just general

thanks, i got V2.1 general installed now, seems to be working. Altough not as fast as i would like. Would you suggest using main stream or just sub stream data on high model (Nvidia P2000 deepstack gpu windows bi machine)

 

[MikeLud1]

thanks, i got V2.1 general installed now, seems to be working. Altough not as fast as i would like. Would you suggest using main stream or just sub stream data on high model (Nvidia P2000 deepstack gpu windows bi machine)

sub stream data on high