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

MicahJames

Pulling my weight
Joined
Aug 26, 2019
Messages
186
Reaction score
212
Location
San Jose, ca
I installed deep stack and enabled it in BI. looks great but when I went to inspect the jpeg used for detection it’s from the sub steam so the quality is terrible. Is there a way to have it scope the main stream ?
 

austwhite

Getting the hang of it
Joined
Jun 7, 2020
Messages
93
Reaction score
92
Location
USA
I installed deep stack and enabled it in BI. looks great but when I went to inspect the jpeg used for detection it’s from the sub steam so the quality is terrible. Is there a way to have it scope the main stream ?
When it is just Object (person or otherwise) detection, it doesn't need to be high res. Even Sentry uses the substream. It would take a long time to process if you sent a high res jpeg to the AI, which would make it useless for triggering motion.
Facial recognition may need higher resolution, but BI Integration and AI Tool does not yet do that.
You can set BI to record a High Res JPEG in settings on trigger if I recall, but have not played with that setting.
As long as DeepStack or whatever AI you use can make out if it is a person or not, resolution doesn't need to be higher.

As a note, if you want to send high res to the AI, which is not recommended due to speed of processing, you would have to turn the substream off, which is also not recommended.
 

Ssayer

BIT Beta Team
Joined
Jan 5, 2016
Messages
19,582
Reaction score
70,740
Location
SE Michigan USA
Like others, I have found that DeepStack is fantastic during the day, but rejects way too many things at night. I'm guessing that changing the mode to high, which BI currently is not giving the option for, might help. But... what do I know? :p

Anyway, in the meantime, to most easily have the best of both worlds, under AI for each cam, I have unticked the "Hide cancelled alerts on timeline and all alerts AND ticked the Auto-flag confirmed alerts. That way clicking on the Flag shows me only confirmed DeepStack alerts, but clicking on the Lightning Bolt shows me everything.

I'm sure that, with time, I won't need to do that. For now though, this way works best (for me)...
 

austwhite

Getting the hang of it
Joined
Jun 7, 2020
Messages
93
Reaction score
92
Location
USA
Unfortunately object detection on any system is pretty hit and miss on black and white images lit by infa-red. Even some on camera person detection fails badly in that situation.
I think the only way around it is to use scheduled profiles to use standard motion detection at night and AI during the day.
That said. I have one Reolink RLC-822A outside and Deepstack actually does detect people and cars very well from that even at night. Go figure that. The crappiest camera out there, and Deepstack works decently even on it's night vision....... Not that the footage is usable as you can't ID people from the crap footage, but it is enough for DeepStack to detect people from it.
 

sebastiantombs

Known around here
Joined
Dec 28, 2019
Messages
11,511
Reaction score
27,690
Location
New Jersey
I'm glad it's not just me seeing problems with B&W. I saw that early yesterday evening and went in and tinkered with brightness and contrast on one camera but it didn't help so I shut off DeepStack in it and in one other camera. Everything was fine. When I checked alerts from overnight that camera stopped detecting anything just after midnight. No schedule change or any other change, I was examining the inside of my eyelids at midnight. I think it needs some work and some fine tuning on the user end for it to work as well at night as it does with full color. Sure, there are more cameras around that are night-color capable, but the vast majority of situations require IR at night.

@Ssayer That's a great tip with alert settings. I'm going to give that a try tonight and see what happens.

When I checked the log, I could see the triggers on that camera that stopped working and they were cancelled by DeepStack even though I had shut it off on that camera. I did reboot the machine at that point, this morning, and everything seems fine at least for now. I'll see how it goes with a full day of bright sun versus tonight.
 

Dan111

n3wb
Joined
Mar 15, 2021
Messages
27
Reaction score
9
Location
19382
Like others, I have found that DeepStack is fantastic during the day, but rejects way too many things at night. I'm guessing that changing the mode to high, which BI currently is not giving the option for, might help. But... what do I know? :p

Anyway, in the meantime, to most easily have the best of both worlds, under AI for each cam, I have unticked the "Hide cancelled alerts on timeline and all alerts AND ticked the Auto-flag confirmed alerts. That way clicking on the Flag shows me only confirmed DeepStack alerts, but clicking on the Lightning Bolt shows me everything.

I'm sure that, with time, I won't need to do that. For now though, this way works best (for me)...
I notice for me it detects people as cars sometimes and cars as people. Not sure if there's a way to change that? It just detected a fox as a person lol. I'm using deep stack off of a gpu.
 

MicahJames

Pulling my weight
Joined
Aug 26, 2019
Messages
186
Reaction score
212
Location
San Jose, ca
When it is just Object (person or otherwise) detection, it doesn't need to be high res. Even Sentry uses the substream. It would take a long time to process if you sent a high res jpeg to the AI, which would make it useless for triggering motion.
Facial recognition may need higher resolution, but BI Integration and AI Tool does not yet do that.
You can set BI to record a High Res JPEG in settings on trigger if I recall, but have not played with that setting.
As long as DeepStack or whatever AI you use can make out if it is a person or not, resolution doesn't need to be higher.

As a note, if you want to send high res to the AI, which is not recommended due to speed of processing, you would have to turn the substream off, which is also not recommended.
Thank you for this
 

IAmATeaf

Known around here
Joined
Jan 13, 2019
Messages
3,287
Reaction score
3,252
Location
United Kingdom
I used Deepstack with AITools quite a while back and also found the it worked perfectly during the day but was hit and miss during the night, since then I stopped using it and now rely upon in camera detection which does have more false positives but at least nothing is missed.
 

Dan111

n3wb
Joined
Mar 15, 2021
Messages
27
Reaction score
9
Location
19382
I used Deepstack with AITools quite a while back and also found the it worked perfectly during the day but was hit and miss during the night, since then I stopped using it and now rely upon in camera detection which does have more false positives but at least nothing is missed.
It is interestingg I still have IVS enabled on my dahuas and I can't tell for sure although it appears that the AI will match the IVS alerts or potentially cancel out IVS alerts. I walked outside and there is no record of me walking outside lol whether I check alerts, confirmed alerts or cancelled alerts. I was closer to the camera than some alerts I have seen it make that were significantly further. This morning the trash people came around 6am and this camera is very far from the street although slightly aided by an ir illuminator it picked up both the truck and person through some branches I didn't even see until I looked closer. Then in broad daylight at 9am it didn't even see me.
 

tech101

Known around here
Joined
Mar 30, 2015
Messages
1,472
Reaction score
2,125
Location
SF BayArea, USA
I have had deepstack miss items at night way more than it would in day. In day its more like sometime person is being detected as car and misses it here and there but at night no matter what setting I do it seems a lot be missing and some of the cams are in color.

I am hoping as this deep stack integration is very new right now with BI. Once it does get mature sometime in future we have some troubleshooting tools or way to fine tune this even further. Again my first time using deepstack so don't know if I can do anything else already to help.

I did play with Confidence level. Adjusting the Motion sensor and selecting all the interested area/Zone and stuff. I even dropped the confidence level to 5% and still was missing. Today I have picked out additional 2 images every sec for this one cam to test if it help in anyway.

By the way prior to this I was paying for Sentry AI for maybe 2 year or so now. One thing I can say for sure I have cancelled my Sentry AI now and wont be renewing subscription with them.

As BI DeepStack is same pretty much the same for now, and have potential to get better. When I did email sentry AI in the past my experience with them was they are trying to improve further but for 2 year have not seen much.

1617374099008.png
 

Vettester

Getting comfortable
Joined
Feb 5, 2017
Messages
740
Reaction score
693
I used Deepstack with AITools quite a while back and also found the it worked perfectly during the day but was hit and miss during the night, since then I stopped using it and now rely upon in camera detection which does have more false positives but at least nothing is missed.
One of the advantages of having AI integrated into BI is you can now take full advantage of using profiles and have different AI configurations for day and night.
 

cscoppa

Getting the hang of it
Joined
Dec 14, 2019
Messages
50
Reaction score
26
FYI I had a lot of trouble getting this to work. Somehow I followed the link to download the windows CPU version and got a super old, activation-required version. Burnt 2 hours trying to figure out why it wasn't working.

Solution - download the latest DeepStack windows binary from Releases · johnolafenwa/DeepStack
GPU version from your link has been working for me for a few days now with BI. (Was using AI Tools before)
 

sebastiantombs

Known around here
Joined
Dec 28, 2019
Messages
11,511
Reaction score
27,690
Location
New Jersey
I notice another little "glitch". Fooling this morning I made changes in the AI sections of every camera, 15 out of 17 anyway. AI stopped and motion detection was degraded afterwards and it took a reboot to get things running right again.

@wittaj Another comment regarding using "auto flag" AI confirmed events. Keep in mind that a flagged alert is a protected alert and will effect disk space utilization wherever you're storing them. I shut off "auto flag" as a result. Too many confirmed alerts in a relatively short time will run you out of disk space.
 
Last edited:

fenderman

Staff member
Joined
Mar 9, 2014
Messages
36,897
Reaction score
21,250
I notice another little "glitch". Fooling this morning I made changes in the AI sections of every camera, 15 out of 17 anyway. AI stopped and motion detection was degraded afterwards and it took a reboot to get hing running right again.

@wittaj Another comment regarding using "auto flag" AI confirmed events. Keep in mind that a flagged alert is a protected alert and will effect disk space utilization wherever you're storing them. I shut off "auto flag" as a result. Too many confirmed alerts in a relatively short time will run you out of disk space.
You can disable protect flagged alerts.
 

t_andersen

Young grasshopper
Joined
Oct 5, 2014
Messages
90
Reaction score
17
Location
Sweden
What's the basis for selecting CPU vs GPU?

I'm assuming that if you have a higher-end GPU that would be the way to go.

But I don't.

My BI server is an i7-6500 with a (somewhat pitiful) Radeon R5 340X (2MB). Card isn't used for anything. Server runs at about 14% baseline/20% max CPU depending on what's happening.

Stick with CPU or offload to GPU?
Deepstack is much faster with a GPU than a CPU alone because it can parallelize its task with a GPU. However, you need a really hefty gaming GPU from NVidia for Deepstack, so unless you have that, then stay with the CPU version.

I have used Deepstack for almost a year with one of the implementations described in this forum and I have also seen that it doesn't work well at night with headlights etc. On the other hand, the human eye also has trouble at night.

When it comes to computation time, several posts here claim that computation time does not increase much due to Deepstack. That is both right and wrong. If Blue Iris sees a person and sends it to deepstack, then a single image is really easy for deepstack. However, if there is a spider web in front of the camera or it is snowing heavily, then deepstack gets an image to analyze every second or two, and then the CPU gets very busy. Then my CPU load goes up from 10-15% to maybe 75% for hours and the computer becomes red hot. (I just ordered a Jetson Nano to try to offload the AI to that one, don't know if it will work).

Some people here wrote that the new version of BI does not allow you to limit recordings to those "verified" by Deepstack? Is that really true? Does it mean that I will get hundreds, if not thousands, of clips when a spider decides to climb in front of the camera? That would definitely be bad.
 

cscoppa

Getting the hang of it
Joined
Dec 14, 2019
Messages
50
Reaction score
26
Deepstack is much faster with a GPU than a CPU alone because it can parallelize its task with a GPU. However, you need a really hefty gaming GPU from NVidia for Deepstack, so unless you have that, then stay with the CPU version.

I have used Deepstack for almost a year with one of the implementations described in this forum and I have also seen that it doesn't work well at night with headlights etc. On the other hand, the human eye also has trouble at night.

When it comes to computation time, several posts here claim that computation time does not increase much due to Deepstack. That is both right and wrong. If Blue Iris sees a person and sends it to deepstack, then a single image is really easy for deepstack. However, if there is a spider web in front of the camera or it is snowing heavily, then deepstack gets an image to analyze every second or two, and then the CPU gets very busy. Then my CPU load goes up from 10-15% to maybe 75% for hours and the computer becomes red hot. (I just ordered a Jetson Nano to try to offload the AI to that one, don't know if it will work).

Some people here wrote that the new version of BI does not allow you to limit recordings to those "verified" by Deepstack? Is that really true? Does it mean that I will get hundreds, if not thousands, of clips when a spider decides to climb in front of the camera? That would definitely be bad.
I haven't tried it yet myself, but someone here said you can clone the camera, and in the Alert Actions, put the HTTP URLs in you would use in AI Tools to trigger recording.
 

sebastiantombs

Known around here
Joined
Dec 28, 2019
Messages
11,511
Reaction score
27,690
Location
New Jersey
Yes, the human eye has a problem with headlights at night. On the other hand if it has headlights and is moving it's a dead giveaway that it's a vehicle.

I guess you're recording on alerts only. Yes, BI will only record a verified alert from what I've seen, but I am recording everything 24/7.

From what I can see, the problems with night time AI verification and with missed verifications during the day would seem to indicate that it isn't really ready for prime time just yet. It does catch the overwhelming majority of events during the day and hardly any at night but the one you want/need might just be the one that gets missed.
 
Last edited:

Ssayer

BIT Beta Team
Joined
Jan 5, 2016
Messages
19,582
Reaction score
70,740
Location
SE Michigan USA
Yep. Checking my street cam today... since 2:30pm, 17 vehicles went by and I got alerts from all of them in BI. 14 of them were verified by DS. Looking at colors, size, silhouette, etc., DS caught very similar vehicles to the ones it didn't. I had min confidence % set at 6 the whole time. For the time being, I'm glad I have it set up to record both ways.
 
Top