IReallyLikePizza2
Known around here
Where do you find that? Nothing in main settings, and I can't find any other AI settings
Hell Yeah! Direct Deepstack Integration - 5.4.0 - March 31, 2021
- Thread starter fenderman
- Start date
Where do you find that? Nothing in main settings, and I can't find any other AI settings
Holy smoke, I set mine to 2 and it killed my CPU dead
poor old i5 4th Gen. I need an i7 or maybe i9 
MikeLud1
IPCT Contributor
MikeLud1
IPCT Contributor
Since the CPUs do not have enough power more the one instance will not work. RIP LOLHoly smoke, I set mine to 2 and it killed my CPU dead
IReallyLikePizza2
Known around here
I've switched Rekor back to CPU and will try running 2 instances with lower sample rate in Bi and see how it works out. I have a GPU but
IReallyLikePizza2
Known around here
MikeLud1
IPCT Contributor
IReallyLikePizza2
Known around here
Maybe I read those instructions wrong, I thought it was implying you needed to specify which Deepstack server/instance was being used for each camera if you have more than once instance
This is interesting, I really didn't pay attention to this setting, mine has been on 1 until now. What is the recommended if you have GPU with some reasonable grunt? and an old i5 processor
and 10 cameras mix of 2mp and 4mp. I will of course experiment, as you do 
Question, does having 2 instances instead of 1 with lower + real-time images give better results? equal load on GPU/CPU maybe relatively.
IReallyLikePizza2
Known around here
Connected a Noctua NA-RC6 LNA to the fan on the P4 to make it a little quieter, can't hear it outside the case now and temps are holding steady. It will handle full load with no issue
sebastiantombs
Known around here
My fault. I should have mentioned it probably isn't going to work, increasing instances, with the CPU version due to capability limitations.
I didn't set anything other than raising the instances to two. I'll let BI and DS figure out. They can think faster than I can.
I didn't set anything other than raising the instances to two. I'll let BI and DS figure out. They can think faster than I can.
Last edited:
I will stick with 1, high and my usual 15-25 +real time images
sebastiantombs
Known around here
I'd like to get a sanity check on building a custom model from the group here, at least those who have built their own custom models.
My original thought was to create a custom model for night use using captures from my own video. I've got a problem with lighting and headlight bloom which makes even the dark model struggle and miss quite a few. As I've been making jpgs it occurred to me that I could add a few more things, like deer, raccoon, fox, opossum, cat, dog, person, car, truck, bus and so on. If I do that I could use both day and night captures and create a complete, master, style model that is specific to my location and cameras. I could then use just that model, no dark and not even the stock model.
The problem with that idea, combined, master, model is that it becomes a big investment in time and effort due to the sheer number of jpgs needed to cover all the objects I want to capture. My target is at least 30 to 50 jpgs for each object, day and night. Multiply that by the number of objects and it's getting into the 300 to 500 range already and could conceivably get bigger depending on how many objects I settle on.
Am I nuts doing this? Is 30 to 50 jpgs enough for each object or should I shoot for more? Am I spinning my wheels?
My original thought was to create a custom model for night use using captures from my own video. I've got a problem with lighting and headlight bloom which makes even the dark model struggle and miss quite a few. As I've been making jpgs it occurred to me that I could add a few more things, like deer, raccoon, fox, opossum, cat, dog, person, car, truck, bus and so on. If I do that I could use both day and night captures and create a complete, master, style model that is specific to my location and cameras. I could then use just that model, no dark and not even the stock model.
The problem with that idea, combined, master, model is that it becomes a big investment in time and effort due to the sheer number of jpgs needed to cover all the objects I want to capture. My target is at least 30 to 50 jpgs for each object, day and night. Multiply that by the number of objects and it's getting into the 300 to 500 range already and could conceivably get bigger depending on how many objects I settle on.
Am I nuts doing this? Is 30 to 50 jpgs enough for each object or should I shoot for more? Am I spinning my wheels?
Just to know what is your GPU idle temp?
Maybe a change of paste and clean on mine as it is not a new model, is ti?
Mine is around 57/58c idling.
GTX 970 Strix series
I'm guessing flow issues - only 1 small fan at the back - I'm going to install a couple of exhaust fans on the top.
I've taken the glass side off for now.
I know the fans don't kick in until 60 so nothing critical I think.
We've had hot weather this week also.
51c with an extra spare USB fan blowing in there. Looks like I need some exhaust fans on the top
Last edited:
I think you're thinking is just fine, agree though a master model is a big job, hence why I went with night only and around 50 images from existing footage, it works well for headlight bloom, taillights etc.
Honestly if I could have just one model covering all it would be good but potentially that is a lot of work. If that one model improved performance then it would be a bonus, for now I just run default during daylight hours and both during the night. I may well have another go at training my night model with some additional images, when I used to use just motion and no Ai some cameras would catch wildlife, maybe some training for deer, foxes would be useful. Hardly essential for security but I like to see what is roaming about out in the wild
As a side note, some of my cameras are worst case scenario, no artificial light, near to a low volume road and trying to monitor both property and passing traffic. This is a big challenge, IR only and high shutter speeds, compromise really. On the plus side training a model for headlight bloom is pretty easy, just need to be careful regarding moths and insects etc triggering false alarms.
Last edited:
That would be a good call, looks like you have an intake running already. No sense wasting those fan mounting points