Deepstack false detections

Does anyone find a solution for this? I am constantly getting notified for no reason. However, when there is an actual incident it doesn’t always notify.

also this one, using deepstack with iSpy

I have put about 5 million photos through DeepStack with my 3 security cameras (yes, literally). I have gotten about 5000 cases where I think that there were legitimate high confidence miss identifications (completely false) objects/artifacts as people. A lot of this is weather dependent. Just today I had “people” show up a few times that were a result of streaks of very large snow flakes close to the camera. It does happen. If you were expecting perfection wait 5 to 20 years.

I have rarely (maybe 100?) events where people were missed. This typically isn’t a problem for me since I process photos a minimum of once per second during motion. If I miss someone on one photo the chances of getting them on another are very, very high. I’d define an event as a series of photos where people were moving through the camera view.

All things considered, this isn’t half bad.

I do have my own software (On Guard) that allows me a lot of flexibility in masking in/out areas of interest. That helps a lot. If you are interested in trying it I’ll provide a link.

The dog identified as a person in your sample photo looks like it is probably the “fault” of the very long shadow. BTW, with On Guard the dog could/would/should very probably not trigger an alert due to the fact that you can define minimum object height on a per area basis with a defined overlap to that area. A person at that location would probably be much taller than the dog even with the shadow. Proper area definition(s) would qualify the confidence as high enough, but the height as too low, thus avoiding the false. It isn’t automatic, but it can be done with experience.

Again, as long as you don’t expect perfection DeepStack is very useful. You need to expect a period of adjustment in changing parameters in third party software as well. You’ll need to adjust how many pictures are taken at what intervals with what degrees of confidence, with what masks. I would expect that with the proper settings you will only very rarely completely miss important activity.

Yet again, do expect false alarms. You can minimize them, but you can’t eliminate them. This is particularly true for outdoor cameras. I would guess that indoors (warehouse, etc.) there would be very few. The rate of false alarms should be very dramatically less that just relying on your camera hardware unless you have a very high end camera.

3 Likes

Very great feedback and comment @Ken98045 . We really appreciate the insight you shared and really hope others who come to read this thread will benefit greatly from it.

@Furkha and @AssAssassin02 , I am sorry to see that you get false and inaccurate detections sometimes with DeepStack. DeepStack standard detection API was trained on the COCO dataset, which has a very broad range of scene images in it but doesn’t cover all other factors like

  • camera angle (mobile vs cctv vs aerial)
  • shadows
  • raining/snowy conditions
  • dark scenes
  • etc.

Reasons like this is why we introduce custom models where users can re-use detections from their specific scene images to train a custom detector and deploy across all platforms/operating system DeepStack support.

Also, we do take the effort to publish custom models for specific scenarios, such as the DeepStack-ExDark we released about a year ago specifically for detecting people, vehicles and a few other objects in dark/night scenes.

To get maximum performance from the camera views generated from your scenes, you can always

1 Like

@Ken98045 @OlafenwaMoses I’m already excited since Deepstack suppressed almost 99% of false alerts, thanks for the suggestion, I just download dark.pt and restarting deepstack now. Thanks.

1 Like