IP Cam Talk Custom Community DeepStack Model
- Thread starter MikeLud1
- Start date
Sybertiger
Known around here
What's the story with the default 300 Epoch runs? I'm sure you are running it all the way through for completeness and accuracy but I was looking at the DS custom models video, which doesn't have narration, and it looks like the author was pointing his mouse at one of the stats of the Epoch runs to indicate that enough accuracy was achieved. He didn't allow it to run all 300 and simply grabbed the best.pt file fairly quickly as in "it's good enough" for gubermint work. Just asking because I was creating a model for for chits and giggles and let it run all 300 Epochs....was thinking I may could have grabbed the best.pt file after 50 runs.
MikeLud1
IPCT Contributor
You can try the best.pt after 50 runs, I did notice that the best.pt file is larger then the final best file. My original models I ran 60 Epochs. I am not sure if that effects the model. What yolov5 model are you using to train yolov5s, yolov5m, yolov5l, or yolov5x. All of my models I am using yolov5s this will have the fastest detection times.
Sybertiger
Known around here
I was using the default yolo5m on the Google Colaboratory website since I don't have a GPU card in my box. I was just fooling around with creating a USPS model specifically for my cam's field of view. The USPS logo model (created by some other fellow) works pretty good but I was trying to create a "uni-directional" model to capture only incoming not outgoing mailman (dead-end neighborhood). I was seeing if I could use my existing IVS triggers that are bi-directional. I just might create a clone cam then use zones to create a uni-directional trigger then use the USPS logo model.....don't know, just fooling around.
Last edited:
MikeLud1
IPCT Contributor
I just watched the video and the training in the video was done with with yolov5m which should have 42Mb file, but when he show how the model works the file size was 14Mb so this model was trained with yolov5s.
MikeLud1
IPCT Contributor
Update: General V2.1 is in first post
ipcamtalk.com
General V2.0 is finished. I am not happy with the results (I started training V2.1 hopefully it will turnout better). I tried something different with V2.0 , tried to balance the number of image from the original general model and the dark model because the dark images have less images then the original general model.
The model does work, I am using it right now for all my cameras. So I attached the model if anyone wants to try it, there is a pdf in the zip file with details of the model.
General V2.0 Results:
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...
General V2.0 is finished. I am not happy with the results (I started training V2.1 hopefully it will turnout better). I tried something different with V2.0 , tried to balance the number of image from the original general model and the dark model because the dark images have less images then the original general model.
The model does work, I am using it right now for all my cameras. So I attached the model if anyone wants to try it, there is a pdf in the zip file with details of the model.
General V2.0 Results:
Last edited:
MikeLud1
IPCT Contributor
I started training general V2.1. This model is being trained with 82,825 images (12.7GB) and will take 5 days to train. When I trained V2.0 I used 20,235 images (3.51GB) and took 1.5 days to train .
Last edited:
sebastiantombs
Known around here
@MikeLud1 I ran the general, V2.0, model overnight on my two problem child cameras. It actually performed quite well. The west camera detected about five more times versus the dark model and the east camera got them all. So far daytime is working at 100%. I've enabled it on another, indoor, camera to see how it does with people. Detection times are very consistent between day and night which is a nice thing.
MikeLud1
IPCT Contributor
Which dark model were you comparing to, mine or original?
sebastiantombs
Known around here
I was running both your and the original simultaneously. Yours was labeled "dark" and the original was "dark1". I know the original is substantially larger and to throw in another variable I raised the detection from medium to high, but the overall performance has been quite good. Detection times are under 200ms in all cases.
Charles_CO
Young grasshopper
Does it make sense to run both combined and dark models if I have lights turn ON with motion at night?
Just curious how will dark model work when lights will turn on and camera will switch from IR mode to regular mode with enough ambient light from the lights.
Just curious how will dark model work when lights will turn on and camera will switch from IR mode to regular mode with enough ambient light from the lights.
sebastiantombs
Known around here
I'm not a fan of motion lights. Most miscreants don't even pay attention to them. That said, running dark and combined gives you the best of both worlds. Dark for before the lights come on and combined once they're on.
Is anyone else using the new "general" model? If so how's it working for you?
Is anyone else using the new "general" model? If so how's it working for you?
Charles_CO
Young grasshopper
@sebastiantombs
I use dumb lights with smart switches and turn on lights around the house when we get home after sunset or when we hang out on the deck when it's dark. I use cameras as a trigger to tell home automation to get the lights going. Using deepstack help with false triggers but it take a little bit more time to trigger the lights. I am just playing with it to see what works, what is practical etc. With enough light at night I would think we could get better camera footage as well.
Is general model good for both day and night?
I use dumb lights with smart switches and turn on lights around the house when we get home after sunset or when we hang out on the deck when it's dark. I use cameras as a trigger to tell home automation to get the lights going. Using deepstack help with false triggers but it take a little bit more time to trigger the lights. I am just playing with it to see what works, what is practical etc. With enough light at night I would think we could get better camera footage as well.
Is general model good for both day and night?
sebastiantombs
Known around here
Yes, the general model works for both day and night.
Given that even the creator of this new model (@MikeLud1 ) wasnt very happy with it, I wanted to wait it out a few more days until he has the next version available (this coming weekend perhaps?).
Your experience has been helpful to read though, and I am somewhat tempted to try this version based on your satisfaction with it.
sebastiantombs
Known around here
I am getting some new falses with the general model, but not enough to be concerning. Otherwise it has been doing a nice job and I'll be switching all cameras over to it later this evening.
I'm usually not one that rushes to the latest beta version or even the latest release. In this case I figured that Mike went through a lot of trouble and work to put this together so the least I can do is give it a shot to see if it works and, hopefully, my observations help with the next iteration.
I'm usually not one that rushes to the latest beta version or even the latest release. In this case I figured that Mike went through a lot of trouble and work to put this together so the least I can do is give it a shot to see if it works and, hopefully, my observations help with the next iteration.
MikeLud1
IPCT Contributor
I am using general 2.0 on all my cameras and it is works, but I think I can do better. See the stats below, higher is better
General V1.2
General V2.0
Charles_CO
Young grasshopper
I will do some testing for general 2.0 as well! I will still have to use animal model too as I really need to "keep an eye" on deer! That said so far I did not have much luck with deer. Most of them were recognized as sheep, dog or cow.. LOL But I am sure it's part of my learning journey.
MikeLud1
IPCT Contributor
Yes