Frigate+ Base Model 2024.1 Update

[blakeblackshear]

NOTE: The base model was updated 24 hours ago. If you recently requested a new model, check the base model version before requesting a new one.

The base model in Frigate+ has now been updated to 2024.1. All model requests going forward will use this version. In order to get a new model, you will need to submit a model request.

In addition to incorporating new images from Frigate+ users, this base model update was focused on reducing known false positives. To date, nearly 250,000 false positive examples have been submitted by users. The previous base model reproduces 26% of these false positives and the new base model reduces that to 11%.

I also updated the fine tuning process to increase the effect that your verified images have on your model. My fine tuned model now reproduces only 1.5% of my reported false positives.

Here are some stats showing the false positive reduction for my model.

Model	Results
2024.0 Base Model	Reproduces 18.6% of my reported false positives
My Tuned 2024.0 Model	Reproduces 16.7% of my reported false positives
2024.1 Base Model	Reproduces 9% of my reported false positives
My Tuned 2024.1 Model	Reproduces 1.5% of my reported false positives

According to every objective measure I have looked at, this should be a significant improvement in false positive rates, but I expect results to vary by user. I would appreciate some user feedback (both good and bad) after upgrading.

[Whomever6378]

Very excited to try the new model, there has been one plant in particular out the front of my home that supposedly looks exceptionally human 😅
I had actually just submitted a new model request maybe 5 hours ago?
Will this happen to be on the new 2024.1 or should I expect to submit another request for the new base model?

[blakeblackshear]

Any model requests submitted in the past 24 hours will be using the new model.

[YBonline]

Although ideally the AI should figure it out, if its one plant in particular, you can mask that object on that area. I have multiple tables with umbrellas, and I've had to mask those spots because Frigate is very insistent the umbrellas are people with extremely high confidence even after training tons of photos..

[Benoulimou]

Hello ,
Thank you for all your work on Frigate. I use it all the time as a complement to other things while I refine it and I really hope to succeed in using it exclusively!
However, is it possible to add an option for multiple deletion of images sent to Frigate+ in order to redo models more in line with recent settings (change of cam position for example). It would make things a lot easier.
Oh yes, and is it possible to delete models?

Sorry for the questions, thank you very much for all your work! I can't wait to try it out!

[NickM-27]

If you move a camera you shouldn't be deleting old images, I've done this multiple times without any issues

[Benoulimou]

There are really a lot of them in the old position, won't that "disrupt" the model?

[NickM-27]

No it shouldn't. I've moved my garage camera twice and even took down a camera that had a few hundred images and it hasn't caused any issues at all. Blake has also answered that it shouldn't cause any issues in most cases.

[Benoulimou]

thank you for your reply

[RWMoo]

Very excited to try this, really appreciate the hard work you put into this. I'll provide feedback over the next week :)

[bbrodriguez]

I’ve been eagerly awaiting this, 2024.0 was great and set a high bar.

Just got my 2024.1 completed and configured in and super keen to see how it fares.

edit: I’d love to see a write up / summary of what changes you made to training sample selection this time.

[blakeblackshear]

Very similar to the last round, but I weighted the sampling to focus primarily on false positive examples.

[bbrodriguez]

My initial impression is that it is an improvement. I dropped my thresholds a little based on the stated reduction in false +ve’s - to try to increase detections through the night.

Of the images below, I might have expected at least one cat detection to have been labelled a dog under 2024.0.

Admittedly it was a clear night so not challenging images to process.

It still seems strong for people at night also which is important to me.

[bbrodriguez]

So a few false +ve detections creeping in, but I can bump my thresholds back up a little (dialled them back to 0.75) which would eliminate many.

What I do think would help is improved motion detection (which I understand is in the works) - specifically tracking an object through its motion. I seem to suffer occasional bouts of a table being detected as a dog while a dog is in the frame somewhere - maybe after it passes by.

[NickM-27]

specifically tracking an object through its motion

frigate already does this, a false positive dog and the real dog are most likely unrelated unless you see them as the same event when using the recordings timeline viewer

[blacknell]

I'm wondering how we could get false positive statistics for our own data - like you summarised in your intro.

[blacknell]

Excellent

[blacknell]

If it's relatively straightforward would be good to get a view of how many false/true positives/negatives there are by label/camera so I can focus on adding those to frigate+ ready for next model request. with 400+ images it's getting difficult/becoming guesswork as to where to focus

[NickM-27]

I don't know that the goal should be even label distribution. I think the goal should be focusing on where defections are missing or boxes are inaccurate

[blacknell]

Fair point but it would show gaps by label

[pyrodex]

I wrote a simple Python script that does this for my needs, it allows me to submit true/false by label, camera, zone, or a combination of all three. It also allows me to purge events based off those factors if I ever decide to do so.

~$ frigate-events.py -h
usage: frigate-events.py [-h] [--silent] [-y] [-p] [-s] [-n] [-f] [-t] [-l LABEL] [-c CAMERA] [-z ZONE]

Utility script to help process Frigate events in large amounts

options:
  -h, --help            show this help message and exit
  --silent              Disable verbose output
  -y, --yes             Enable automatic YES instead of being prompted
  -p, --purge           Purge events
  -s, --submitted       Process submitted events only
  -n, --not-submitted   Process not submitted events only
  -f, --false-positive  Submit the events as False Positive to Frigate+
  -t, --true-positive   Submit the events as False Positive to Frigate+
  -l LABEL, --label LABEL
                        Specify Frigate Label, these are always lower case (e.g. person)
  -c CAMERA, --camera CAMERA
                        Specify Frigate Camera, these are the same CASE as displayed in your Frigate configuration
  -z ZONE, --zone ZONE  Specify Frigate Zone, these are the same CASE as displayed in your Frigate configuration

[mintjberry]

Honestly I was very underwhelmed by my first model using 2024.0, even after tuning it wasn't a lot better.
Am excited and hopeful for 2024.1, have requested my first model using it, thanks.

[rough316]

Tried my first + model, its not working. I get this in the logs:

2024-04-26 03:19:54.587111853 Process detector:ov:
2024-04-26 03:19:54.587125103 Traceback (most recent call last):
2024-04-26 03:19:54.587134163 File "/usr/lib/python3.9/multiprocessing/process.py", line 315, in _bootstrap
2024-04-26 03:19:54.587142563 self.run()
2024-04-26 03:19:54.587150983 File "/usr/lib/python3.9/multiprocessing/process.py", line 108, in run
2024-04-26 03:19:54.587159223 self._target(*self._args, **self._kwargs)
2024-04-26 03:19:54.587171553 File "/opt/frigate/frigate/object_detection.py", line 125, in run_detector
2024-04-26 03:19:54.587179833 detections = object_detector.detect_raw(input_frame)
2024-04-26 03:19:54.587196023 File "/opt/frigate/frigate/object_detection.py", line 75, in detect_raw
2024-04-26 03:19:54.587204733 return self.detect_api.detect_raw(tensor_input=tensor_input)
2024-04-26 03:19:54.587212953 File "/opt/frigate/frigate/detectors/plugins/openvino.py", line 84, in detect_raw
2024-04-26 03:19:54.587233863 infer_request.infer([tensor_input])
2024-04-26 03:19:54.587242323 File "/usr/local/lib/python3.9/dist-packages/openvino/runtime/ie_api.py", line 152, in infer
2024-04-26 03:19:54.587251023 return super().infer(normalize_inputs(self, dict(enumerate(inputs))))
2024-04-26 03:19:54.587259413 RuntimeError: The input blob size is not equal to the network input size: got 307200 expecting 270000

[NickM-27]

Process detector:ov:

You're trying to use openvino which isn't currently supported

[rough316]

Was thinking thats the case, trying to get it working with CPU, then will go find my USB Coral.

[rough316]

Up and running on the Coral USB. You guys are doing a fantastic job on Frigate!

[Sch0field9]

I have statues of lions on my gate pillars, which are often mistaken for either a person or a dog. (Interestingly, they are almost never mistaken for a cat.) From the camera's perspective, people almost always walk behind the statues.

You mentioned that to reduce false positives, we should submit true positives in that area along with the false positives. Now, if a person walks behind one of these statues and I submit it as a true positive with the statue overlapping the person, would that confuse the model?

[Sch0field9]

Regarding that topic: the camera in question is a PTZ camera, making it impossible to mask out the statues. Is there a solution to this issue, or is something planned for zones/masks on PTZ cameras?

[blakeblackshear]

As long as the person is actually visible too, it should be fine.

[blacknell]

So in this circumstance where the Frigate+ identified box is too big should we mark this not a package and draw a tighter bounding box?

[YBonline]

I believe you say it is a package, and then adjust the size of the box before verifying

[bbrodriguez]

I’ve noticed the face boxes are extremely large in 2024.1.

[NickM-27]

Yes, this would be this is a package and fix the box.

Face boxes look accurate just like always in my usage so far

[YBonline]

So my initial results are that false positives are indeed way down, but false negatives are up as well (which is obviously a bit of a bigger deal). This might be because I hiked up my min score on the previous model, so I'm gonna try to reduce that and see how it goes.

The camera with the most notable negative impact in my sole newer (to me) fisheye camera.... its detecting near nothing now. I don't have a whole lot of training on it either, so I'll have to see if I can drop the minimums pretty low and start training it. I do have it using the raw stream, with no de-warping, so I don't know if that is a proper goal or if there should be a different model for this unusual camera format?

I just didn't want to create multiple additional streams as I have so many cameras already, and I'm near the limit of what my hardware can support without more RAM.

[nshp]

Looking pretty good so far with the new model. Still some false positives and misidentifications here and there, but the false positives have been noticeably reduced, making it much easier to scroll back through events. The only identification that doesn't seem to have been reduced is deer, but most of those are relatively low confidence, I'll just set the minimum higher. The stationary object tracking seems to be working much better for me with parked cars. It was losing the car-ness of parked cars many times a day previously, now it's pretty solid. I frequently have two cars that overlap from the camera's perspective, I could watch the detection going nuts hopping back and forth between which one of them was a car and which wasn't.

I did notice one particular type of new false positive showing up. One of my cameras can see the street through some (sparse) tree branches. I've been submitting plenty of true positives of cars driving by seen through the branches, Frigate has been doing a fine job of catching those. Now it seems I've trained this new model that cars, and branches blowing in the wind, are cars. I'm submitting those false positives now, hopefully it'll go back to normal with the next model I train.

I think my biggest remaining problem model-wise is that squirrels and foxes are always running by firing off a confident stream of "cat - dog - cat - dog - cat - dog" events.

[NickM-27]

Some of these sound like things that the new UI will greatly improve the experience of.

Stationary tracking has also been improved for the next version.

[krustophenia]

Thanks for your excellent work!! One question, I've submitted some images with dog labels from different cameras, day/night etc but my frigate is not detecting any since I started using this custom models. Should I keep uploading and labeling dog images? I've lowered the score for dogs to 40-50% and even like this no detections. Thanks.

[NickM-27]

You should include your config, seems like a config issue

[krustophenia]

model:
  path: plus://XXXXXXX
mqtt:
  host: XXXXXXX
  user: XXXXX
  password: XXXXX
detectors:
  coral:
    type: edgetpu
    device: usb
cameras:
  Cam_Piscina:
    ffmpeg:
      inputs:
        - path: rtsp://XXXXX:XXXXX@XXXXXX:554/cam/realmonitor?channel=1&subtype=0
          roles:
            - detect
        - path: rtsp://XXXXX:XXXXX@XXXXXX:554/cam/realmonitor?channel=1&subtype=0
          roles:
            - record
            - rtmp
    motion:
      improve_contrast: true
  Cam_Passadis:
    ffmpeg:
      inputs:
        - path: rtsp://XXXXX@XXXXXX:554/cam/realmonitor?channel=1&subtype=0
          roles:
            - detect
        - path: rtsp://XXXXX@XXXXXX:554/cam/realmonitor?channel=1&subtype=0
          roles:
            - record
            - rtmp
    motion:
      improve_contrast: true
      mask: 
        - 389,203,402,0,1280,0,1280,576
  Cam_Parking:
    ffmpeg:
      inputs:
        - path: rtsp://XXXXXX@XXXXXX:554/cam/realmonitor?channel=1&subtype=0
          roles:
            - detect
        - path: rtsp://XXXXX@XXXXXX:554/cam/realmonitor?channel=1&subtype=0
          roles:
            - record
            - rtmp
    motion:
      improve_contrast: true
  Cam_Frontal:
    ffmpeg:
      inputs:
        - path: rtsp://XXXXXX@XXXXXX:554/cam/realmonitor?channel=1&subtype=0
          roles:
            - detect
        - path: rtsp://XXXXXX@XXXXXX:554/cam/realmonitor?channel=1&subtype=0
          roles:
            - record
            - rtmp
    motion:
      improve_contrast: true
  Cam_Menjador:
    detect:
      width: 1920
      height: 1088
    ffmpeg:
      output_args:
        record: -f segment -segment_time 10 -segment_format mp4 -reset_timestamps 1 -strftime 1 -c:v copy -c:a aac
      inputs:
        - path: rtsp://XXXXXX/live0
          roles:
            - detect
            - record
    motion:
      improve_contrast: true
  Cam_Carrer:
    detect:
      width: 640
      height: 480
    rtmp:
      enabled: False 
    ffmpeg:
      input_args:
        - -avoid_negative_ts
        - make_zero
        - -fflags
        - nobuffer+genpts+discardcorrupt
        - -flags
        - low_delay
        - -strict
        - experimental
        - -analyzeduration
        - 1000M
        - -probesize
        - 1000M
      output_args:
        record: -f segment -segment_time 10 -segment_format mp4 -reset_timestamps 1 -strftime 1 -c:v copy -c:a aac
      inputs:
        - path: rtsp://XXXXXXX@XXXXXX:554/h264Preview_01_sub
          roles:
            - detect
            - record
    motion:
      improve_contrast: true
    objects:
      filters:
        dog:
          min_score: .6
          threshold: .7
        person:
          min_score: .70
          threshold: .75

detect:
  width: 1280
  height: 960
  fps: 3
  enabled: True
  max_disappeared: 25

record:
  enabled: True
  events:
    pre_capture: 5
    post_capture: 10
    retain:
      default: 10

snapshots:
  enabled: true
  timestamp: true
  bounding_box: false
  retain:
    default: 10

timestamp_style:
  position: "br"
  format: "%d/%m/%Y %H:%M:%S"
  color:
    red: 0
    green: 0
    blue: 0
  thickness: 1
  effect: "solid"

birdseye:
  enabled: True
  width: 1280
  height: 720
  quality: 1
  mode: motion

objects:
  filters:
    dog:
      min_score: .4
      threshold: .5
    person:
      min_score: .80
      threshold: .82  

[NickM-27]

You haven't added dog as an object for frigate to track (and you also haven't added face, losing a feature of frigate+ model as well)

[krustophenia]

oh right, my bad. Adding this and let's see if that improves. Thanks!

[csreynolds]

At risk of resurrecting a horse to beat; was a list of labels for future models firmed up?

I'm keeping my raised vegetable gardens squirrel free with homeassistant, a smart plug and powered water valve to spray them if frigate detects them in frame.

Currently the squirrels are detected as birds, so I am hesitant to update the model without a squirrel tag or something to keep this working as it is currently.

Not a complaint in any way, this project is great and all your hard work is very appreciated. 🥇

[nwish]

No changes to the detectable objects that I can see.

[blacknell]

I'm seeing (subjectively) a huge drop in false postives particularly at night. I'm also seeing a small number of new false positives (new in new places). So I guess this is 10 steps forward and 1 back. Brilliant.

[OnFreund]

Same here

[dz4news]

Same here... I have actually dropped the min_score on a few objects and still have very few false positives.

[OnFreund]

I'm thinking of lowering min_score too, to avoid cases where one logical event is split into many shorter ones. Not sure if that's needed as much with 0.14 though.

[krustophenia]

Regarding face detection, I've added face as object:track and added

face:
min_score: .80
threshold: .82

shouldn't that do the trick?

[NickM-27]

You won't get face events, just that face will be labeled and used to choose the best image for person

[miziu]

Hello, is the release date of the version 2024.2 known?

[blakeblackshear]

Updates will be roughly every 3 months.

[rotilho]

Anyone noticed a reduction in inference speed? I just moved away from the OV detector to the Coral detector. The inference with the original model with the Coral detector is 9ms, but with the Plus 2024.1, it's around 140ms.

Is it expected? I remember reading somewhere that the Plus models would be more lightweight.

[mimainka]

For me - using the Coral as well - the inference speed stayed practically the same after switching to Plus. I'm always between 7 and 13 ms.

[NickM-27]

For me it is faster, from 8 Ms to 5ms

140 ms is problematic and indicates an issue

[blakeblackshear]

Probably means you grabbed the CPU model ID instead of the edgetpu ID.

[rotilho]

You are totally right. I got the wrong model!

[hagak]

Can you run different models against different cameras on the same frigate instance? Reason I ask is the Frigate+ Model does not have "Boat" and I have one camera that faces water and like to get notified when a boat is going by. I get that is not a focused object but if i could run the base model on that one camera and then the Frigate+ on all my others would be great.

[NickM-27]

no, this is not possible

[hagak]

figured it wasnt. :(

Will boats be part of the Frigate+ anytime soon? :)

[NickM-27]

expanding the label set is coming up soon on the roadmap https://github.com/orgs/frigate-nvr/projects/2/views/1

[f1d094]

@blakeblackshear: If you were to guess, when would you be generating Frigate+ Base Model 2024.2 update?

[blakeblackshear]

Late July

[bbrodriguez]

Any update available on ETA?

[f1d094]

Is there a place to provide feedback and/or track wishes, hopes, and desires for new object support? Where we live we have a strong desire to know more about wildlife on the property. Notably:

• Racoons
• Skunks
• Mountain Lions
• Coyotes
• Bears*

All of the above all track as other supported objects but some of the above would be really nice to be able to track independently for obvious safety reasons. Or nuisance reasons, as far as the trash pandas. Maybe I'm alone on this but it might be worth tracking who needs support for what.

Also: Why no USPS ? Err..."post-office"...more geographically independent, I suppose.

*We theoretically have bears around. I have never seen one.

[blakeblackshear]

I'm working to add bears and some other animals for the July update (2024.2) assuming I can find enough examples and the model doesn't get confused.

Also have plenty of USPS examples.

The challenge isn't in adding the new labels, but in everything that surrounds it. Once you get to 20+ labels, it's really hard to keep all the things you need to look for in the image in your head. A lot of things need to be reworked to support a much larger set of labels including workflow, suggestions for already verified images, changing verified to be per label, etc. Also, just because something can be labeled doesn't mean the model can support it. This needs to be clear.

Once I get the base training set back filled with the new labels I am trying to add, then I can train and evaluate if they can even reliability detect. I won't know which new labels I will be able to add until that happens.

[dlarkinsPonderosa]

What is the best way to supply snapshots of things we can't currently label? Bears.

I have bears all over the place. They are typically captured as cats. I do not flag those as false positives for Frigate+ as I feel an event of some kind is better than no event. (I don't upload them as verified cats though).

Should I upload those images and label nothing? I'm also concerned that might train my model that there are no cats to be seen. (Which there aren't of course, but I want my cat bears).

Happy to help provide examples.

[NickM-27]

for now probably best to just not submit, once the bear label is added that should be submitted

[blakeblackshear]

Actually, I would upload them without labeling. If the images have no labels, they don't even get used during training, so they won't impact anything. These are such clear examples that my first iteration of a model with bear label will probably pick them up and then I can grab them.

[f1d094]

@blakeblackshear Semi-non-sequitur to my prior question: What is the best practice with regard to model-training for unsupported objects that are tracking as supported objects? Of the list I mentioned, they each track to similar-but-incorrect supported objects:

• Racoon -> Cat
• Skunk -> Cat
• Mountain Lion -> Dog or Cat or Deer
• Coyote -> Dog
• Bear -> N/A

I do want to track the above and have no idea how many we are missing out on. The percentage confidence on these varies wildly. Currently, any image with any of the above I simply ignore in hopes that someday the above will be supported and I don't want to pollute the model. I presume this is best-practice.

Please advise.

[blakeblackshear]

I would suggest either not labeling them or labeling as something they are visually indistinguishable from. I'm not entirely sure if it will ever be possible to differentiate between some animals reliability. It may ultimately take a secondary pass with a slower classification model or something.

[tbclark3]

I also want to track raccoon (a nightly occurrence) and coyote (two in the last week, believe it or not, in downtown Tacoma WA). Both are currently detected as "cat" and I have stopped submitting "not cat" to plus for the same reason as @f1d094 . I eventually want my model to detect coyotes and raccoons, although I hope I have no need of bears. It would be really helpful to know what objects are in the works and when they are likely to be included. It would also be helpful to know what the proper action is now with regard to objects not currently detected. Do we say "Not cat" or do nothing?

Also, @f1d094, as a workaround, I am using the Google generative AI integration to resolve all the "cat" questions. It returns the correct animal almost 100% of the time, including fuzzy nighttime images with IR light.

[NickM-27]

I think not submitting them is the best path right now, that is what I do for raccoons

[blacknell]

Foxes and deer in the UK are principle nuisances

[tbclark3]

Would it be possible to add filtering to the Frigate+ images page? Following your most recent advice, I would like to edit all of the "not cat" images but with >1400 images, it isn't practical to scroll through them. I can get the most recent ones from Frigate images Edit in Frigate+, but most of the ones that I want to edit have long since been removed. Simple filtering based on the presence of a label or "not label" would be adequate. This issue is likely to become more pressing as more labels are added, particularly if it is done in stages.

[blakeblackshear]

Totally possible. I have a pretty long list of enhancements and this is on there.

[tbclark3]

Also, could you clarify your instructions about submitting animals without a label? What if there is another known object, such as a person or car, in the same image, and those are labeled but the animal is not? I'm guessing that with at least one unrelated label, the image will be included in the model. Won't that teach the model that the unlabeled coyote shouldn't be recognized?

[blakeblackshear]

It may or may not include that part of the image depending on where they are. Hard to say exactly how much of an impact that would have in reality.

[derekcentrico]

Not to be that guy, but I have a couple new cameras and a lot of new photos in + waiting to build a model. Don't want to burn a run with the likelihood of 2024.2 coming around soon based on prior comments. Is it still on target for this month?

[blakeblackshear]

It's going to take longer this time since I am trying to add new labels, so I don't think it will be ready in July. I may decide to release a new model without the new labels if I think it's going to take more than a few more weeks to get there.

[derekcentrico]

It's going to take longer this time since I am trying to add new labels, so I don't think it will be ready in July. I may decide to release a new model without the new labels if I think it's going to take more than a few more weeks to get there.

I vote to take your time and make that happen. I'll burn another build from the quota in the meantime!

[j-vaughn]

It's going to take longer this time since I am trying to add new labels, so I don't think it will be ready in July. I may decide to release a new model without the new labels if I think it's going to take more than a few more weeks to get there.

@blakeblackshear Is there a way to buy more credits before our subscription renews? I didn't see an option in the portal. I'm just starting my year and am trying to decide if I should do my second run now to handle some false positives or wait for the new base.

[blakeblackshear]

I haven't added that functionality yet, but I can help you figure it out if you reach out to me via email at [email protected].

[blacknell]

Is there anyway to see how many images were used on various model builds. For example, my last model was trained in April (on base 2024.1) but I’m not really sure how many new photos I uploaded and labelled since then. Regards Paul

…

On 22 Jul 2024, at 00:29, derekcentrico ***@***.***> wrote: It's going to take longer this time since I am trying to add new labels, so I don't think it will be ready in July. I may decide to release a new model without the new labels if I think it's going to take more than a few more weeks to get there. I vote to take your time and make that happen. I'll burn another build from the quota in the meantime! — Reply to this email directly, view it on GitHub <#11106 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACDT75HVNEADWPSYIHTKUQ3ZNQ75JAVCNFSM6AAAAABGYURZECVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTAMJQHEYDSMY>. You are receiving this because you commented.

[OnFreund]

You can compare the number of verified images you have today (in the dashboard) with the number of images used in the previous model (in the models screen).

[dabibbit]

I have a different take on the other requests people made for more animal labels: If more animal labels are added, adding labels for livestock and predators would be especially helpful. There are a ton of uses for Frigate and HA alerts on a farm once Frigate can make determinations about the animals. Are the cows on the field side of the fence? Awesome. Is there a calf that got out and is now in the zone between the fence and the road? Have HA set off every light and siren in the house to wake me up before the calf dies on the road and hurts someone in a car.

While identifying common livestock is helpful for setting alerts and actions based on them being where they are supposed to be and not being someplace other than expected, identifying common farm predators is useful for obvious reasons. Fox outside the chicken coop? Interesting data. Fox found a way inside the chicken coop? Time is of the essence.

There are pretty good numbers / economics on which farm animals are most common. You can't go wrong with chickens, cattle, sheep, ducks, goats, pigs, horses, donkeys/mules, and llamas. Many of the predators have already been mentioned by others. A good roster would be coyotes, racoons, snakes, foxes, weasels, possums, mountain lions and bears.

My personal priority list would be cattle, goats, sheep, ducks and coyotes. Then add in deer, since they're all over and getting a positive ID as a deer would likely prevent it from being classified as something else. Lumping together cows, bulls, steers and heifers as cattle works for me.

I know @blakeblackshear mentioned earlier that it gets a bit difficult to juggle the labels once there are a lot of them. However, I do think the farm use case could make that worthwhile. Beyond homeowners who are annoyed at certain animals getting into the trash, having a robust and extensive set of animal labels can provide lots of benefits on a farm (both economic and life preserving).

[YBonline]

Just as an FYI, the current model does already detect foxes as dogs pretty reliably. I have a fox trying to get my chickens, so I already have an automation that flashes lights and makes my squeaky motorized gate move a bit and it reliably scares the fox away from them, at least so far.

[The1Percent]

Just curious as I am working on my first data submission to start frigate+, I have two frigate systems using the same cameras with identical configurations both using yolo_nas_s one set at 320 and one 640. The 640 detections seems to be far more accurate especially at any distance. Do the plus models use a 640x640 snip or 320x320?

[NickM-27]

I believe the Frigate+ yolo nas model is trained on 640x640 regions and then the input for the model is set at 320x320, @blakeblackshear can confirm

[blakeblackshear]

That's correct. I didn't see a significant difference in performance after downsizing, but it significantly reduces inference times. I could produce a 640x640 version as well from the training pipeline.

[The1Percent]

It would be super interesting to see if the results are different as I have two identical systems taking streams from the same set of cameras. The inference times are lower on 320 for sure. I didn’t notice a huge hit like when I tried running yolonas_m, but these arc 380s are kind of amazing for the money. Ill finish collecting data for the next few days and them submit. From: Blake Blackshear ***@***.***> Sent: Tuesday, October 22, 2024 5:27 AM To: blakeblackshear/frigate ***@***.***> Cc: The1Percent ***@***.***>; Comment ***@***.***> Subject: Re: [blakeblackshear/frigate] Frigate+ Base Model 2024.1 Update (Discussion #11106) That's correct. I didn't see a significant difference in performance after downsizing, but it significantly reduces inference times. I could produce a 640x640 version as well from the training pipeline. — Reply to this email directly, view it on GitHub <#11106 (reply in thread)> , or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACE353EECBVXMN5LPBMM33TZ4ZACPAVCNFSM6AAAAABGYURZECVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTCMBRGY3DQMA> . You are receiving this because you commented. <https://github.com/notifications/beacon/ACE353AD2JB5P3OD32MLXFTZ4ZACPA5CNFSM6AAAAABGYURZECWGG33NNVSW45C7OR4XAZNRIRUXGY3VONZWS33OINXW23LFNZ2KUY3PNVWWK3TUL5UWJTQAVAM6Q.gif> Message ID: ***@***.*** ***@***.***> >