I see this very often as well. Mine typically doesn’t recover either.
I would suggest most people use 0.5.0 or 0.5.1-rc1 for the time being. The changes I made for rc2 clearly are not ready for general use. Further updates from me will be slow over the next month.
@surge919 I would suggest restarting the container via docker instead.
Hi Blake, thanks, will try that.
I’ve reverted back to 0.5.1-rc1
I’ve tried creating automation to restart the container but haven’t had any luck.
If anyone here has done this, please let me know how.
My coral arrived today. Is it as simple as plugging it in and rebooting and frigate will auto use it or do I need to do something else?
It was for me with my Synology.
I knew it was working because without it, there are messages saying “No EdgeTPU detected”
Also there was a noticeable increase in the detection FPS
1 Like
I’ve just checked my log and it says no edgetpu detected, falling back to cpu. I’m on Ubuntu 18.04 elitedesk. Anyone any ideas?
You need to install the coral drivers, check their website
1 Like
You might not have to install the coral drivers since they are in the container. I didn’t install any drivers on my host. Are you sure you are using all the parameters from the example command or compose file in the readme?
If I remember right I needed a reboot after install
I think thats me up and running guys. thanks for your help. 
What version of Live555 is the Dafang hack using?
The Wyze version is pretty dated
2020-03-24T13:23:06.646414579Z s=Session streamed by "wyze"
2020-03-24T13:23:06.646515630Z i=live
2020-03-24T13:23:06.646560513Z t=0 0
2020-03-24T13:23:06.646623886Z a=tool:LIVE555 Streaming Media v2017.10.28
Transfer learning would be amazing for this, unlocking so many more possibilities, I saw there was an attempt at a PR for saving detections to use later for learning, any chance it can be revisited? Would be awesome to be able to start collecting a dataset early on.
I’m not 100% sure how to find that out. Searching the github I see some v4l2rtspserver options to use live555 (-s), but I don’t think dafang is using it, it’s using -r I believe.
I’m happy to poke around more if you have any hints.
also - searching google I get this result https://gitter.im/Xiaomi-Dafang-Hacks/Lobby?at=5cb2a14a3ebbdc55b3b3d970
I found mention in their issues section in GitHub and it seems like it uses the same version.
The snippet I showed above was from the beginning of the logs after a start up of the container.
I rolled back to rc1 and the behavior is still just as erratic.
edit… I checked your link and that seems to be the same reference mentioned in GitHub
I have seen a few postings on here about f+’s lately. I’m not sure if it was integrating the new models, or perhaps something else, but I am seeing the same thing. I ran the older versions for almost a year and had exactly 1 false positive, and that was because I had accidentally left one of my thresholds at 70%. However, over the last month or two I have had many. Everything from birds, to spider, to dogs. Just today a dog walked by and frigate was 96% certain it was a person. That is a really high degree of certainty. I am using min and max, but unless all your detections are all the same distance from the camera, you can only use them within reason. In the case of the dog, it was a largish dog. Not sure what the cause is, but I wanted to chime in with my experiences. I have HA set to use priority messages to wake my wife and I up in the night for leaks, people skulking around the house, etc. We used to never get false alarms, but now it’s common.
It would be great if you could save a video clip with the false detections and include at least 30 seconds before the false positive. When I start working on reducing the false positives, it will be helpful to have a set of videos to run through for testing.
Sure. All my detections kick off a 15s recording, but I think I only get 5s before detection. I’ll try to change that in HA.
It’s interesting. Seems like there should be far fewer false positives with this implementation because only starts detection when it sees motion and doesn’t have detection on the other 99% of time to get false positives. Also tailoring the detection area accurately to the object (which it does a great job at) should greatly increase accuracy (at least for me who just had 2 big areas defined on each camera). I was getting quite a few false positives in the old version of just parts of my lawn or wall or tables, etc, which I think I won’t see any more. That was about 80% of my problems.
So far I’ve had no false positives on the new version, but I haven’t been running it very long. I can see much higher accuracy when I test just going out and walking around at least, though it does seem to take a second or two so for the motion detection to find me.
Animals will always be an issue, but I’ve filtered most of them out just by size (even though I also have a big range of distances in my camera). I wonder if adding specific categories for cat and dog and bird into your config would help here? It might have a higher detection value for cat than human and then not detect it as human…?
Simple question, I’m on an old version of frigate (~0.2.0) but I’m not sure which, how do I check, in case I need to roll back?