Are you using Low, Medium, or High mode. All my models are trained to use High for the best results, I am seeing better confidence compared to my general model even during the day
IP Cam Talk Custom Community DeepStack Model
- Thread starter MikeLud1
- Start date
Sybertiger
Known around here
i5-8500 CPU version of Deepstack
I'm using the combined (Day) and dark (Night) and thought I was doing good until I noticed my detection times going up. I tried to get my detection times down but I must be in a mental block as my times keep going up...LOL. Take a look and please point out the obvious. And why it decided that the 80% detection was better than the 91% analysis is beyond me.
I'm using the combined (Day) and dark (Night) and thought I was doing good until I noticed my detection times going up. I tried to get my detection times down but I must be in a mental block as my times keep going up...LOL. Take a look and please point out the obvious. And why it decided that the 80% detection was better than the 91% analysis is beyond me.
Attachments
Last edited:
Try rebooting to see if the detection times come down. For the 80% detection was better than the 91% this is Blue Iris that makes the decision.
Sybertiger
Known around here
sebastiantombs
Known around here
@MikeLud1 You're working with color at night. The scenes I have are B&W since there's nowhere nearly enough light for color even on a 5442. The headlight bloom, even with HLC at 100 is quite strong and there's only a frame or two where an actual vehicle shows in the frame without being obscured by the headlights. That's one of the things about models that can be problematic, they are based on nice images under the best conditions and probably shot with an SLR to boot.
Last edited:
Our systems couldn't be a more exact match: same CPU (and I also run the CPU only DeepStack installation), same cameras, same substream settings, same medium setting for photo size on the Global Ai settings page. Yet my analysis times are dramatically lower, on the order of 90-350ms (per photo) generally. The only thing that stands out to me (and I could be wrong) is the syntax that insures that the default model isn't used at all: yes, I noticed you unchecked it in the Global AI dialog box, as did I, but I also include the "objects:0" in each AI trigger dialog. Also, why not just eliminate the references to the dark model on the profiles that won't use it (Daytime) and only use the "combined" model there?
Which version of DeepStack (Sept 2021 or early 2022)? I run the (much maligned) early 2022 version, but haven't had any problems with it so far.
I am not at the console at the moment, but could help later if you have any other comparisons you want to do.
Edit: I see your reboot helped tremendously...we cross posted.
Looks good, now you are getting around GPU timesWelp, I think the reboot helped just a little.
View attachment 129214
View attachment 129215
View attachment 129216
View attachment 129217
I am curious though: why run DeepStack on top of the built-in camera AI triggers? I consider those considerably more reliable than the BI motion detection triggers, and do the opposite: I run both types, but DS analysis only on the conventional motion triggers to filter out some of the noise (shadows, headlights across the parked cars in my driveway, etc..).
kc8tmv
Getting the hang of it
I have noticed BI presenting the lower percentage several times as well.
Sent from my iPhone using Tapatalk
I always thought as soon as BI sees the percentage threshold being reached by Deepstack object detection, it alerts and stops considering other images. This would not be the case if you include a cancel object that doesn't exist in the model.
Sybertiger
Known around here
I would think that BI would build a list (percentages) during the evaluation period then pick the best one and if it meets/exceeds the threshold then it would go with that one. If DS runs 3 evaluations and positively reports a car detection each time (i.e. 65%, 80%, 70%) then I would have thought BI would report the 80%.
I started training General V2.0. The training will take about 36 hours
Wow, what a transformation! Looks like the back will be a full reno. Lawn work is a hobby of mine. I started my back lawn from dirt this past fall. TTTF/KBG mix.
Last edited:
Sybertiger
Known around here
You are probably referring to pics like these. All come from 5442 cams that would not be able to operate in color for my situation. The dark model could not identify these. Typically, I see that black/dark colored vehicles are difficult for the dark model to detect although you can see a white police SUV in one pic was missed, maybe due the the reflective decals. These were all triggered by IVS in the 5442 cams for human/vehicle so we know the cams AI can detect them.
sebastiantombs
Known around here
You're lucky with broadside shots and I would think that with a little tinkering DS using combined and dark would pick those up quite easily. The ones I get are much worse than those and I do get detections about 5% on one camera, cars west and 95%, or better, on cars east. Cars East has an advantage from a streetlight being in the scene providing better illumination.
