Hell Yeah! Direct Deepstack Integration - 5.4.0 - March 31, 2021

sebastiantombs said:
@105437 I could swear that you need to copy some files from the cuDNN subdirectories into the CUDA directory structure but I can't remember where I saw that. I'll look in my office later, I think I printed out the directions and that will have a URL on it.
Click to expand...
I found the cuDNN install instructions and moved the files where they needed to go. Things are working pretty well. Below is pretty typical processing times. I'm running the standard Objects model and the Dark model. I'm only running one instance of DS, should I be running more than one now that I have a GPU?

0 11/19/2021 5:02:07.561 PM Driveway_Cam DeepStack: car:93% [343,57 626,182] 328ms
0 11/19/2021 5:02:07.561 PM Driveway_Cam DeepStack: Car:92% [329,54 634,190] 328ms
 
  • Like
Reactions: sebastiantombs
You should be able to bump the instances up. Go up one at a time and watch detection times with each increase. They should get shorter then jump up pretty high when the capabilities of the card are exceeded.
 
  • Like
Reactions: 105437
Let us know how it works out.
 
  • Like
Reactions: 105437
kklee said:
Nvidia driver of your video card.
The CUDA Toolkit 10.1
The CUDA Toolkit 10.1 update 1
The CUDA Toolkit 10.1 update 2
Extract the cuDNN NVIDIA lib into your CUDA Toolkit folder.
Click to expand...
I'm awaiting a P400 in the mail next week. Can these steps be performed in advance of the hardware being installed (to get a headstart on the ordeal)?
 
MikeLud1 said:
Split it up by cameras that might use DeepStack at the same time.
Click to expand...
I'm running two DS instances now and I just had a FedEx delivery. Overall, I think the times are a little worse than just running with once instance. The Street, Driveway and Front Porch cams are on the same instance. The Entry, Front and Doorbell cams are on the other.

Here are the events from the delivery... Thoughts/suggestions would be appreciated!

3 11/20/2021 10:29:32.962 AM Street_Cam MOTION_A
0 11/20/2021 10:29:39.466 AM Street_Cam DeepStack: Car:86% [51,67 166,156] 575ms
0 11/20/2021 10:29:39.466 AM Street_Cam DeepStack: car:79% [56,75 161,152] 575ms
3 11/20/2021 10:29:50.902 AM Street_Cam MOTION_A
0 11/20/2021 10:29:57.409 AM Street_Cam DeepStack: Car:88% [102,109 265,265] 342ms
0 11/20/2021 10:29:57.410 AM Street_Cam DeepStack: truck:75% [111,119 254,256] 342ms
3 11/20/2021 10:30:04.769 AM Entry_Cam MOTION_A
0 11/20/2021 10:30:11.239 AM Entry_Cam DeepStack: Car:95% [336,100 705,407] 330ms
0 11/20/2021 10:30:11.239 AM Entry_Cam DeepStack: car:88% [343,109 692,395] 330ms
3 11/20/2021 10:30:19.970 AM Front_Cam MOTION_A
3 11/20/2021 10:30:22.020 AM Front_Porch_Cam MOTION_A
3 11/20/2021 10:30:25.425 AM Driveway_Cam MOTION_A
0 11/20/2021 10:30:28.107 AM Front_Cam DeepStack: truck:86% [128,25 364,127] 1426ms
0 11/20/2021 10:30:28.107 AM Front_Cam DeepStack: Car:86% [116,13 374,139] 1426ms
0 11/20/2021 10:30:30.595 AM Front_Porch_Cam DeepStack: Car:86% [510,3 692,69] 1137ms
0 11/20/2021 10:30:30.595 AM Front_Porch_Cam DeepStack: car:74% [521,0 691,62] 1137ms
0 11/20/2021 10:30:32.341 AM Driveway_Cam DeepStack: Car:80% [465,18 658,174] 2298ms
0 11/20/2021 10:30:32.341 AM Driveway_Cam DeepStack: truck:73% [473,30 652,167] 2298ms
3 11/20/2021 10:30:54.619 AM Driveway_Cam MOTION_A
3 11/20/2021 10:30:59.880 AM Front_Porch_Cam MOTION_A
0 11/20/2021 10:31:01.719 AM Driveway_Cam DeepStack: person:81% [744,139 783,214] 342ms
0 11/20/2021 10:31:01.720 AM Driveway_Cam DeepStack: truck:80% [471,31 651,165] 342ms
3 11/20/2021 10:31:05.028 AM Doorbell_Cam MOTION_A
0 11/20/2021 10:31:07.550 AM Front_Porch_Cam DeepStack: People:88% [612,28 680,162] 935ms
0 11/20/2021 10:31:07.550 AM Front_Porch_Cam DeepStack: person:86% [620,32 674,154] 935ms
0 11/20/2021 10:31:07.551 AM Front_Porch_Cam DeepStack: Car:78% [488,3 578,73] 935ms
0 11/20/2021 10:31:07.551 AM Front_Porch_Cam DeepStack: car:71% [494,4 572,67] 935ms
0 11/20/2021 10:31:11.642 AM Doorbell_Cam DeepStack: People:93% [452,170 629,569] 384ms
0 11/20/2021 10:31:11.643 AM Doorbell_Cam DeepStack: person:91% [461,170 618,568] 384ms
3 11/20/2021 10:31:20.086 AM Driveway_Cam MOTION_A
0 11/20/2021 10:31:26.593 AM Driveway_Cam DeepStack: Car:86% [466,20 659,174] 337ms
3 11/20/2021 10:32:07.962 AM Front_Cam MOTION_A
3 11/20/2021 10:32:09.534 AM Entry_Cam MOTION_A
3 11/20/2021 10:32:09.879 AM Front_Porch_Cam MOTION_A
3 11/20/2021 10:32:14.375 AM Street_Cam MOTION_A
0 11/20/2021 10:32:18.940 AM Front_Cam DeepStack: Car:86% [106,15 355,138] 628ms
0 11/20/2021 10:32:18.940 AM Front_Cam DeepStack: truck:84% [114,21 346,127] 628ms
0 11/20/2021 10:32:20.597 AM Front_Porch_Cam DeepStack: car:73% [663,2 753,48] 1662ms
0 11/20/2021 10:32:20.598 AM Front_Porch_Cam DeepStack: Car:71% [650,0 761,54] 1662ms
0 11/20/2021 10:32:21.820 AM Entry_Cam DeepStack: Car:92% [294,113 525,315] 3685ms
0 11/20/2021 10:32:21.821 AM Entry_Cam DeepStack: car:85% [308,119 514,307] 3685ms
0 11/20/2021 10:32:24.169 AM Street_Cam DeepStack: truck:93% [281,41 854,476] 5575ms
0 11/20/2021 10:32:24.170 AM Street_Cam DeepStack: Bus:89% [274,24 856,468] 5575ms
 
Hey Hey,

I hope this is the right place to ask these questions:
  1. I would like to trigger recording and an alert only when deepstack is able to recognize the object multiple times for a period of a few seconds.
    The idea is to prevent the detection of cars or persons that pass by "quickly". In other words, I am only interested in people that linger around.

    I assumed this can be configured using the option "Wait until triggered at least" in the alerts tab. Unfortunately, I still get alerts (push notifications) that contain "nothing found" as detected object but without any recording now.
    What am I missing? :)
  2. Is there an option to not detect static objects or objects that have been detected previously?
    E.g. our parked car is always detected when a person passes by and I would like to only detect the person.
  3. Can you train deepstack or correct it? I had a few funny detections (a person being detected as car)

Thank you!
 
@Otanaught - in Blue Iris, simply make the make time a little longer so that the quickly passing objects don't trigger. Depending on your field of view, you could up that to 1 second or more.

In the AI setup tab, you can uncheck the detect static objects option

Yes you can train your own custom model. You use your pictures from your field of view. Others here have said it is like an hour or so to train a custom model.
 
  • Like
Reactions: Otanaught and sebastiantombs
Thank you @wittaj. Finding the right settings and amount is tough in that vast array of options.
Is there maybe also a setting to handle motion detection when it rains? Right now there is light rain and the motion detection sensing and triggering is going wild :)

Edit: Even after increasing the make time, I still get a lot of alerts/pushes without any recording. I have the feeling I am missing a setting that would prevent sending an alert when there is nothing to be recorded.
 
Last edited:
Otanaught said:
Can you train deepstack or correct it? I had a few funny detections (a person being detected as car)
Click to expand...

Yes, you can, but no way it takes an hour.
You need to collect and label at least 100 images for each class you want to detect from your CCTV feed (and you have to train the model using pictures of people in different positions, crouching etc...), then your first model will give you a lot of false positives during rain (or snow), then you have to collect that false positive images, include them in your dataset as background class and do another training run (rinse and repeat). That does take a lot of time, plus there will be technical problems - DeepStack documentation asks you to install the latest version of PyTorch which is not compatible with DS, labelimg tool can be buggy, you can accidentally mislabel a couple of images and wonder why training is failing and so on. Own model is a bliss - it can be fast and precise but unless you have some technical background and prepared to invest a lot of time in it I would recommend using default model for daytime detection an exdark model for night time (IR).
 
  • Like
Reactions: actran, Otanaught and sebastiantombs
Thanks @ivanerr the topic just got a lower priority on my bucket list ;)

Does anyone know what the "mark as vehicle" setting does? I assumed that all these objects will be classified, flagged and highlighted as vehicles but it seems to do nothing for me.
 
Otanaught said:
Thanks @ivanerr the topic just got a lower priority on my bucket list ;)

Does anyone know what the "mark as vehicle" setting does? I assumed that all these objects will be classified, flagged and highlighted as vehicles but it seems to do nothing for me.
Click to expand...

This is for people that are using Plate Recognizer with a custom plate model so that it only sends pictures with plates to Plate Recognizer.
 
  • Like
Reactions: aesterling, sebastiantombs, Otanaught and 1 other person
You must log in or register to reply here.
Top Bottom