sebastiantombs
Known around here
AI, like DeepStack, running on a PC is not a foolproof system as you have found out. I get occasional mis-identifications as well, dogs or deer as "people", people as "horse or dog", a shed as a "truck" and a flower bed as a "boat". In my case it doesn't happen often, but no system, even when running on a really high horsepower system. Don't expect results like you see in the movies or on television, ain't gonna happen.
Probably that is the case, when I first installed Blue Iris and Deepstack I like it very much and now that I'm trying to tune it I start to find problems, my system for now is very basic installed in HyperV on an Gen4 I7, I address 7 cores to it but struggles with high cpu everytime I open Blue Iris in foreground.
Until I have all the new cameras installed I won't upgrade the hardware and I'm not also not sure if it's hardware because lately if seems to give more false positives, makes little sense to id a 4 legs animal has a person (unless the coder was thinking in dog style...)
Thanks the help
Until I have all the new cameras installed I won't upgrade the hardware and I'm not also not sure if it's hardware because lately if seems to give more false positives, makes little sense to id a 4 legs animal has a person (unless the coder was thinking in dog style...)
Thanks the help
Flintstone61
Known around here
Some people are dogs. Call the MIB.
Field of view, lighting, and camera settings play a big part of this as well.
Some people try to do too much with a field of view because it works in the movies.
Some screenshots of your camera views with the mis-identification can help us troubleshoot your issues, along with screenshots of your settings.
And then as someone else mentioned, sometimes one will get the occasional mis-identification.
Some people try to do too much with a field of view because it works in the movies.
Some screenshots of your camera views with the mis-identification can help us troubleshoot your issues, along with screenshots of your settings.
And then as someone else mentioned, sometimes one will get the occasional mis-identification.
Here are some images at night I have same configuration with black and white option selected, before most of the times they were dogs and cats now they are persons, I want the system to alert when a person enter my house and don't be alerted by the dogs walking around.
Using DeepStack-Installer-CPU-2022.01.1.exe with --VISION-SCENE=True --VISION-DETECTION=True --VISION-FACE=True --MODE High --
Blue Iris now is 5.5.5.13 x64
I get the feed from my analogic Conceptronic recorder at 12.00 fps /1.00 key day and night, cameras are 1080p angles are lower then 90º probably 4mm lens I just realize that now I'm only using mainstream before I used also substream with less false positives don't know if the problem is here or because of more actual Blue Iris release.
Just changed from edge vector to simple to see if helps.
Attachments
Most of us have found the DeepStack struggles with B/W. If you can run color you may have better results.
Uncheck the use mainstream as DS downrezs anyway, so you are not gaining anything.
Since you cropped the views, I have no idea where these identifications are in relation to the whole field of view. The further you get from the center of the field of view, the more chances you have for a mis-identification.
What are your motion detection settings in BI, perhaps you need to further refine those so that they are not triggering for the size of a dog.
The next thing to do is walk around day and night thru these views. Where dark clothing and see what is the lowest % it IDs as a human. If it doesn't drop below 85%, then no reason to run with a confidence of 60% and then start getting dogs.
Finally, some views are simply problematic with the models that are being used. Many here have created their own model with images from their own cameras to further improve the performance of Deepstack for their needs.
Uncheck the use mainstream as DS downrezs anyway, so you are not gaining anything.
Since you cropped the views, I have no idea where these identifications are in relation to the whole field of view. The further you get from the center of the field of view, the more chances you have for a mis-identification.
What are your motion detection settings in BI, perhaps you need to further refine those so that they are not triggering for the size of a dog.
The next thing to do is walk around day and night thru these views. Where dark clothing and see what is the lowest % it IDs as a human. If it doesn't drop below 85%, then no reason to run with a confidence of 60% and then start getting dogs.
Finally, some views are simply problematic with the models that are being used. Many here have created their own model with images from their own cameras to further improve the performance of Deepstack for their needs.
I don´t have much light so for now have to use BW.
I'm going to reactivate sub-stream and uncheck the use of mainstream in AI, I assumed that better quality was going to improve the ID.
Tomorrow I will know the result.
I'm going to reactivate sub-stream and uncheck the use of mainstream in AI, I assumed that better quality was going to improve the ID.
Tomorrow I will know the result.
CrazyAsYou
Getting comfortable
I have/had a similar issue on a few of my cams, my large brown newfoundland often showed as person in low light but only at worst to about 80% - I did a number of tests and also started using dark.pt at night and found real people are always +85% - my fix was to set confidence at 85% - changing my motion size would have limited detection of real people as the dog is almost as large as a person.