jagged oriole
I'm generally happy with my Frigate detections. I send false-positives (spiders, garbage cans) to Frigate+.
Should I worry about sending in accurate results as well if I have no issues with them? In other words, any issue with just training out the false reports and not reinforcing any of the accurate results? (All I care about is people detection, not cars nor anything else)
forest flower
if you don't label the true positives in your images, the model will start to learn that those are not the objects they are
imagine you are showing images to a person to teach them what different objects look like
some of the images have a "person", but its not labeled
so they look at the image while being tested, say "there is a person right there"
because it wasnt labeled, they get their hand slapped and a "wrong! thats not a person"
eventually they stop guessing that its a person because they are confused and arent sure anymore
forest flower
i was referring more to making sure the true postives are also labeled when reporting a false positive
jagged oriole
Ahh ok. In most of those I don't think there are any true positives missed.
forest flower
but yes true positives are more helpful than false positives generally
jagged oriole
False positives for spiderwebs have been great
forest flower
again, imagine you are teaching someone to recognize objects. if all you ever show them is thousands of things that are "not a person". they will still be left saying "ok, thats great, but what does a person actually look like?"
its more useful to show them what a person does look like
since its all pattern recognition
now the base model obviously already has many many examples of these objects
so its not starting from zero
jagged oriole
Ok I'll start using the green button a bit more. Notably where spiders are a problem.
forest flower
yea. i usually try and grab spiderwebs and other common false positives from cameras and pull them into the base model
jagged oriole
I'm a bit worried of using up all my 12 training runs
jagged oriole
I'm still early days I think at trying to clean up the detections but every iteration/training helps for sure.
Not sure how long training takes and on what hardware, would be cool to either allow local training, and/or allow us to "donate" GPU time (with a limit) to you as well.
upbeat forge
Ahh ok. In most of those I don't think there are any true positives missed.
Just to reiterate, you need to label all objects in the image not just use the green / red button.
jagged oriole
Just to reiterate, you need to label all objects in the image not just use the g...
Yes, however, when I asked before.. if I don't care about cars at all, I can just ignore them right?
forest flower
yea
jagged oriole
I live downtown. Cars every 30 seconds 🙂
People 100% in scope and especially on the side of the house, I need that dialed in.
oak basalt
I don’t care about cars on the road in front of my house but I still let frigate detect them, you never know when it might come in handy if someone drunkenly runs into your place
sonic wyvern
I upgraded from 0.13 to 0.14 and now my two reolink cameras are not able to work in frigate. I have a real simple config running in a container on an amd64 box with a coral USB device connected to it. The box has a real old nvidia card in it that's unused. I'm thinking it might be some codec config issue.
Maybe I'm using the wrong container - :latest?
Any recommendations @forest flower ?
upbeat forge
we need to see logs
sonic wyvern
I truncated my logs and restarted the container. - they should be fresh.
upbeat forge
it's because hwaccel is enabled by default
you were not using it previously which is very much not recommended
in this case, it is using iHD driver but presumably you have an older CPU
so you should set an env variable or add
environment_vars:
LIBVA_DRIVER_NAME: i965
to the config so it will work correctly
sonic wyvern
@upbeat forge , thanks for the response. I tried adding it the environment_vars to my frigate config.yaml and still have no frames have been received. Should I be adding an environment variable to my podman invocation instead?
eg: podman run ... -e LIBVA_DRIVER_NAME=i965 ... frigate:latest
sonic wyvern
model name : Intel(R) Core(TM) i7-3630QM CPU @ 2.40GHz
flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx est tm2 ssse3 cx16 xtpr pdcm pcid sse4_1 sse4_2 x2apic popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm cpuid_fault epb pti ssbd ibrs ibpb stibp tpr_shadow flexpriority ept vpid fsgsbase smep erms xsaveopt dtherm ida arat pln pts vnmi md_clear flush_l1d
vmx flags : vnmi preemption_timer invvpid ept_x_only flexpriority tsc_offset vtpr mtf vapic ept vpid unrestricted_guest
I do run my frigate container with -p 1984:1984. When I point a browser at the go2rtc page. I am able to click the stream buttons and see the camera feeds.
Moar logs after adding the i965 variable.