I still have this agent branch that I have kept separate. I think adding proper semantic search with vectors of each altert image would be huge, but we need a model to do that, and depending on large paid models seems kind of conterintuitive to the whole goal of having self hosted surveilance.
The start I had for semantic search requires passing nearly every detection every time. This is obviously super inefficient. I've been keeping a close eye on Unifi's products and see a need for additional classification within detections, even if delayed after the fact.
OCR accuracy of the ALPR is honestly a separate issue and is improving hugely. What would really be a game-changer is a self-hosted image/video classification that could run in the background and add metadata to all traffic in order to support later research like Flock does.
It's a tough thing to build around because it requires users to have some more powerful localized models or to use public LLMs, but we keep getting better and better open-source models.
I'd really like to see this become a self-sufficient system, and have all functionality for traffic and faces, but the accuracy of the CV is currently a material obstacle. This app isn't meant to step in for any of that, but having more comprehensive data would enable a lot more functionality.
Currently, beyond the pretty simple CRUD functionality, I'm not sure I can really crank up the value much more than what is provided by blue iris and CPAI. I'd like to see more, and I have my own agenda for alert integrations, but I think a lot of this comes back to more trustworthy CV (to use a very academic term) - and getting more from the AI that we can then use for management and research purposes.
The start I had for semantic search requires passing nearly every detection every time. This is obviously super inefficient. I've been keeping a close eye on Unifi's products and see a need for additional classification within detections, even if delayed after the fact.
OCR accuracy of the ALPR is honestly a separate issue and is improving hugely. What would really be a game-changer is a self-hosted image/video classification that could run in the background and add metadata to all traffic in order to support later research like Flock does.
It's a tough thing to build around because it requires users to have some more powerful localized models or to use public LLMs, but we keep getting better and better open-source models.
I'd really like to see this become a self-sufficient system, and have all functionality for traffic and faces, but the accuracy of the CV is currently a material obstacle. This app isn't meant to step in for any of that, but having more comprehensive data would enable a lot more functionality.
Currently, beyond the pretty simple CRUD functionality, I'm not sure I can really crank up the value much more than what is provided by blue iris and CPAI. I'd like to see more, and I have my own agenda for alert integrations, but I think a lot of this comes back to more trustworthy CV (to use a very academic term) - and getting more from the AI that we can then use for management and research purposes.
Last edited:
