Hacker News Reader

Comment by not_that_d | Hacker News Reader

The title is misleading, you get that "full local" just with the default install and model. And the default model sucks, for example this is a "person alert": https://imgur.com/a/uDCCTjr, this is just funny https://imgur.com/a/pP4ZvQI

I have a love/hate relation with Frigate, I use it since 2 years but since the business model of the developer is provide a "good model" using a custom one is not possible (at least not in a easy ways AFAIK).

I use my cameras to track a family member with a medical condition, this is why I do not feel confortable uploading those image to the "Frigate+" service to eventually get better training.

View on HN

I'm probably being dense here, but could you please explain your first image link? Assuming you didn't accidentally link the wrong image, I'm struggling to see how even a not-very-good model would think that's a person, and it doesn't look like a security camera screenshot like your second image does.

Is not the wrong image. I got a notification (thru Home Assistant) because Frigate detected that group of toy horses as a person. I tried to find the image where the box is visible but only found that preview.

Edit: I found a better place to make an screenshot https://imgur.com/a/5qpDWia There you can see the event marked as "Person"

Frigate, Doubletake + Compreface is 100% local

Person you're replying to mentioned Frigate+ which is the paid subscription option offering the ability to upload images to their servers in order to further train the models to get better accuracy, so no longer 100% local.

Maybe you're suggesting that using two additional tools in combination with the free version of Frigate brings its quality up on par with that of an extra-trained Frigate+? If that's so it would be great if you could say that and elaborate how so / why, rather than just dropping in some new tool names and no explanation as to how/if they address GP's points. (Thanks in advance if you do come back and explain.)

Edit: I just looked into Doubletake + Compreface, seems they're both facial recognition tools, so using them wouldn't overcome the problem GP commenter reported that Frigate without Frigate+'s additional training doesn't do a good enough job of general object tagging for them?

Reply on news.ycombinator.com