Cats and dogs recognized as people

[Locutus73]

I'm very pleased of the smart notifications feature, and I use it extensively with an integration which sets off various smart sirens in my home when an unknown person notification email is received from Arlo, but...

...recently my cats and neighbour's dogs are recognized as people very often.

Ok, I love my cats as if they were family people, but... can people recognition be improved?

I donate to Arlo each and every video with a wrong recognition tag, but the issue seems to get worse every day.

Thank you in advance.

Regards.

[ShayneS]

Hi @Locutus73

  What troubleshooting steps have you tried so far? I do not want to suggest anything you may have previously performed.

[Locutus73]

---

@ShayneS  ha scritto:

Hi @Locutus73

  What troubleshooting steps have you tried so far? I do not want to suggest anything you may have previously performed.

---

Hi @ShayneS I tried the usual cam/basestation reboot, deactivating/activating smart notifications and stuff... the problem is erratic, so it's not a glitch in my configuration, it's just that the AI algorithm sometimes confuses cats and dogs for people.

Relocating cameras is not an option: all my 6 cams are in the perfect spots for their intended security use.

Just for your info I noticed that I see this problem happening more frequently with cameras positioned high (approx 2.5 meters) pointing down (probably cameras positioned lower capture animals legs and tails better), and this happens more frequently during night (with all my smart sirens triggered by your email for my pleasure).

As a workaround I alleviated the problem using cloud zones, trying to put my cats passages out of them... but still sometimes I get a random "person" notification during daytime with a small dog clearly visible with legs, tail and all...

Thank you in advance.

Regards.

[ShayneS]

HI @Locutus73

Have you tried removing the cameras from the Active smart list then re-add them? the last step I would suggest before contacting Customer Support would be to remove the cameras from Settings > My Devices > re-add them. Arlo Tech Support

[Locutus73]

---

@ShayneS  ha scritto:

HI @Locutus73

Have you tried removing the cameras from the Active smart list then re-add them? the last step I would suggest before contacting Customer Support would be to remove the cameras from Settings > My Devices > re-add them. Arlo Tech Support

---

@ShayneS the remove+add dance (btw, I tried to remove+add from the Smart plan) won't solve an AI recognition inaccuracy: I haven't a problem related to a bad configuration, i.e. not receiving notifications at all or receiving unwanted notifications... it's just that the artificial intelligence image analysis is flacky and sometimes confuses an animal for a person and no remove+add routine will solve that. Calling the support is useless... they would just tell me to remove, add, reboot, repeat... the development team should just analyze the donated videos and refine the AI.

Thank you.

Regards.

[StephenB]

---

@Locutus73 wrote:
the development team should just analyze the donated videos and refine the AI.

---

Hopefully they are.  Though there is quite a bit of work involved in curating/tagging the donated videos and retraining.

It'd be cool if the AI could be individualized (so I could say "I want ones like this", and "I don't want ones like that".

[Locutus73]

@StephenB I'm glad Arlo staff is supposedly (I hope) tagging and using donated videos to improve AI... but in the meantime I still continue having random videos with cats or dogs recognized as people with consequentely my alarm triggering (not nice at night).

I don't think individualizing AI would help... a cat is a cat, a dog is a dog, I don't need to say "I want ones like this"... Arlo devs can analyze and train the AI with generic cats and dogs videos, not needing my cats and my neighbours dogs...

Thank you.

[StephenB]

---

@Locutus73 wrote:

I still continue having random videos with cats or dogs recognized as people with consequentely my alarm triggering (not nice at night).

---

There will always be some false detections, and of course they do need to balance those with missed detections.  These algorithms aren't likely to improve quickly - it takes a lot of effort to develop a curated training set, and full retraining of an image classification system requires a lot of computation/time.

---

@Locutus73 wrote:

I don't think individualizing AI would help...  I don't need to say "I want ones like this"

---

I think individualing would help, because people often interested in seeing different things - but of course there is room for more than one view on that. 

Even if it's not individualized, the algorithm could potentially improve more quickly if there was direct user feedback.  That would be a form of continual learning - likely a different approach to training than Arlo is using now. Though proper training of an AI algorithm is a complicated topic, and I'm not an expert in that area.