Deepstack Detecting Cats as People (70%+ Confidence)

[IReallyLikePizza2]

Anyone have any good ideas how to solve this? Cameras do not have IVS

Doesn't seem to be working well if it thinks a cat is 73% probable human!

Would increasing the time



Here is how deep stack is setup

 

[sebastiantombs]

All I can say is that not every system is perfect. DS identifies me as a dog sometimes and a shadow on the front yard as a boat. Then again DS is free so I'm not complaining too much.

 

[Fubduck]

If you add cat to list does it identify them correctly?

 

[dz93]

This is common for me too. I'm in the process of training my own model from my own images I collected from my cameras hoping that this can help with more accurate detections. For your situation I might try to increase the min object size under the motion sensor settings. Making that larger can help stop it from detecting the motion from the cats and therefore stop sending cat images to deepstack as often.

 

[IReallyLikePizza2]

Thanks, that helped a lot. Don't know why I didn't do that to start with!

 

[sorka]

Is it worthwhile training models for each camera location with landscape features and such?

 

[icpilot]

This is common for me too. I'm in the process of training my own model from my own images I collected from my cameras hoping that this can help with more accurate detections. For your situation I might try to increase the min object size under the motion sensor settings. Making that larger can help stop it from detecting the motion from the cats and therefore stop sending cat images to deepstack as often.

Here is my problem with using Min Size Target Filters. It starts with the fact I want to get a trigger and alert as early as possible when a person enters the Intrusion Zone or crosses a Tripwire. For my use case, I have a 4mp varifocal zoomed to its maximum setting of 12mm. This provides me a viewing range from about 5 ft in front of the camera out to about 150 ft away from the camera. Ideally, I want a trigger/alert at 120 ft from the camera, so that is where the IVS Intrusion Zone extends as well as Tripwires set. Cats come into view and cause false positives at between 5 ft to 20 ft from the camera. If I set a Min Size to prevent cat alerts at 10 ft from the camera, it's a Min Size of more than 1,000 pixels. OTOH, a 6 ft tall person at 100 ft away is no more than 700 pixels. The Min Size to prevent detection of cats also prevents early detection of potential intruders. While true that as the person comes closer they will (or should) be detected, I've lost the 'edge' of earlier detection.

At this point due to IVS rules being unreliable, I now use IVS along with SMD for Triggers, and Deepstack for Alerts. It works, so I am hopeful AI will continue to improve in eliminating false positives, and also eliminating many of these 'work-arounds' necessary for the current state of AI.

 

[Skispcs]

I am seeing similar instances where things are misidentified.

 

[IReallyLikePizza2]

Maybe set 2 motion zones, one further out with the min object size smaller and one close with a big object

There is no way a buy would be that big out there, or it would be bigfoot

The annoying thing is that the Ring person detection is almost perfect, they have have some secret sauce in there (Or a big building of slaves manually pushing the alert button...)

 

[icpilot]

Maybe set 2 motion zones, one further out with the min object size smaller and one close with a big object

There is no way a buy would be that big out there, or it would be bigfoot

The annoying thing is that the Ring person detection is almost perfect, they have have some secret sauce in there (Or a big building of slaves manually pushing the alert button...)

You're right, that could be done (zones at different ranges with appropriate Min Size boxes). Between the Dahua (or Hik) on camera controls in combination with Blue Iris, the configuration options are plentiful, so workarounds are available. As you point out, the best case would be (will be) a reliable AI algorithm.

 

[sorka]

Maybe set 2 motion zones, one further out with the min object size smaller and one close with a big object

There is no way a buy would be that big out there, or it would be bigfoot

The annoying thing is that the Ring person detection is almost perfect, they have have some secret sauce in there (Or a big building of slaves manually pushing the alert button...)

Their images are sent back to be processed in systems far smarter with more power than anything we are going to have in our on prem PCs.

 

[sorka]

FYI, I have 14 cameras with many of them with waving trees and plants. I have had no false positives and all of my testing hasn't missed a single actual person detection.....of course I've only had it up for two days