What's a good Deepstack processing time?

105437

BIT Beta Team
Jun 8, 2015
2,145
1,046
I'm running Deepstack on the same PC as BI. When I review the log, I'm seeing average response times for DS processing around 2,500ms. Is that typical?
 
I get on average 600ms set on medium, my CPU is an AMD Ryzen 9 3950X and all my cameras are using hardware decode.
Thanks for the reply. I have a 9-year-old PC which probably explains my DS processing time.
 
I've seen times as low as 54ms but they are generally in below 200ms. I'm running an older i-6700K but using the GPU version of DS on an NVidia GTX970.
 
  • Like
Reactions: MikeLud1
I should have added that when I was running the CPU version detection times ranged from ~250 all the way up to 2000+ milliseconds and wasn't very reliable. It has rarely missed using the GPU version.
 
  • Like
Reactions: 105437 and MikeLud1
I should have added that when I was running the CPU version detection times ranged from ~250 all the way up to 2000+ milliseconds and wasn't very reliable. It has rarely missed using the GPU version.
Kind of like this?

0 7/15/2021 9:45:43.321 PM Entry_Cam DeepStack: car:92% [195,115 1304,785] 344ms
0 7/15/2021 9:45:51.512 PM Front_Cam DeepStack: car:89% [143,116 625,289] 1885ms
0 7/15/2021 9:45:53.698 PM Front_Porch_Cam DeepStack: Alert cancelled [nothing found] 2583ms)
0 7/15/2021 9:45:56.296 PM Driveway_Cam DeepStack: car:90% [597,122 1143,425] 2129ms
0 7/15/2021 9:46:01.145 PM Side_Porch_Cam DeepStack: Alert cancelled [nothing found] 743ms)
0 7/15/2021 9:46:09.075 PM Garage_Cam DeepStack: car:91% [930,104 1919,687] 360ms
 
  • Like
Reactions: sebastiantombs
@105437 Yup, that's the kind of performance I saw.
 
  • Like
Reactions: 105437
You'd need an NVidia card specifically. Right now they're pretty expensive due to chip shortages and bitcoin style mining. The mining aspect is dropping off, though, so used cards may come down considerably in the near future. The GT970 uses about 10-15 watts running with DS processing on it.
 
Last edited:
  • Like
Reactions: 105437
I’m not sure a single number is a useful for comparison. As I’ll show below, it hides a lot detail. We‘re probably better off using a lower bound.

For apples-to-apples comparisons I think you need to specify not only the DS version (cpu/gpu), but also if you’ve enabled/disabled these settings - for each camera - you are reporting:
  1. Default object detection
  2. Facial recognition
  3. Custom models; if so, which ones?
On my I7-4770 with only #1 enabled, I’m running around 450-500 ms.

Some of you might be interested in the following charts I‘ve made by extracting data from the Blue Iris logfile for ONE of my cameras over the past 4-5 days. In this example, all data points are CONFIRMED. Deepstack is disabled at night for this camera.

1626485551776.jpeg

1626485569401.jpeg

This analysis could be performed using Microsoft Excel. I’ve automated it using a custom PHP script and the Google Charts API.