Frigate+ 2025.3 Base Model Update

[blakeblackshear]

The base model in Frigate+ has now been updated to 2025.3. All model requests going forward will use this version. If you have a recent model request, it has automatically been updated to use 2025.3.

In this release, sampling was weighted towards labels with fewer examples and you should see significant improvements for bear, deer, fox, kangaroo, motorcycle, rabbit, raccoon, and skunk. Specifically for YOLOv9, I incorporated a more robust sampling of reported false positives into the base model training set.

I have also added many new candidate labels requested over the past few months. These include: rickshaw, wombat, auspost, aramex, bobcat, mustelid, transoflex, airplane, drone, mountain_lion, crocodile, turkey, baby_stroller, monkey, coyote, porcupine, parcelforce, sheep, snake, helicopter, lizard, duck, hermes, cargus, fan_courier, and sameday.

[adamoutler]

I've been using sheep for over a year now. This is apparently a sheep.

[hawkeye217]

That's not surprising. You've been using sheep with the free model that ships with Frigate, based on the COCO dataset. This model is not trained on security camera images. You can explore the images that it's trained on here: https://cocodataset.org/#explore

[lalaricca-ai]

Hey Team. I want to propose a very important new label. "Flashlight/Laser masking" Its really important as Frigate only relies on image data and could be easily tampered.

[zonyl]

I like the idea in theory, however, not sure how practical it will be. Maybe for indoor cameras it will work out ok, but most of my road facing cameras would like trigger this constantly at night from traffic.

I have two different angles of the critical areas around my house for this very reason. I figure it would be very hard for someone to casually blind both locations.

In addition I have a laser perimeter light break that compliments frigate

[NickM-27]

It also doesn't make sense as an object detection model. You'll be able to train classification models in 0.17 and this will work better like that

[zonyl]

Thank you for adding those newer candidate labels! Finally have a label for my otter family (mustelid) and giving family warning of coyotes. I have a sky cam, live near ATL airport approach and a lake so hoping to add some annotations for airplane / drone.

Im seeing some significant improvements in false positives already this past evening.

BTW: A badger is technically a mustelid

[blakeblackshear]

So is a skunk. I'm still narrowing in on when the visual uniqueness within a family warrant a separate label.

[percynodeal]

Big thanks to the whole Frigate crew - devs, contributors, testers etc and the whole community.

So awesome to see the new base image drop with AUSPOST labels. I’ve been wanting to add a label the postie for quite some time now, so I’m stoked to finally start adding some Aussie-specific data into the mix.

Cheers from Australia 🇦🇺

[DavidValeri]

My wife and small dogs thank you for improving false positives on deer. They were tired of being hit with the sprinkler automation I use to keep the deer away from my landscaping.

[zonyl]

Lol. Same here. Sadly we have both deer and stray dog problems where I live. For a while on Yolo9 it was detecting anyone bending over as a dog. Many times we were gardening and getting hit. Needless to say my wife was never amused at being labelled a dog regardless.

[Caros2017]

Maybe it's an "feature request", but are you considering that these new base models will be available with a one time payment instead of a year Frigate+ subscription?

I would love to test these models, but have no intention of using other frigate+ functions, since I just don't like to share images online :)

Currently I am using Yolov9 model s with 640*640 and really wondering how much difference this trained model is.

[NickM-27]

If you purchase a year and then disable renewing then that gets you 4 total base models (this one and then the next 3 releases)

I don't work on frigate+ so I can't say whether exactly what you're suggesting will be available.

[Caros2017]

Thanks for your reaction! :)

@blakeblackshear would you consider this as an option in the near future?

[blakeblackshear]

The one time purchase price is $50. You can cancel your subscription immediately after purchasing to prevent it from renewing. You will have access to all the base models released before your 1 year subscription expires.

[jeeftor]

Or you can just support Frigate+ because it's AWESOME! :) (obviously all people have different financial status)

[baudneo]

Great job Frigate team. Excited to see what the future brings, a la user built classification flows and the triggers.

[wonkydog]

Maybe one for somewhere else,

The new 2025.3 Base Model doesn't appear in my Frigate+ account.

Does this take a while to propagate down and become available to all accounts?

Thanks- Looking forward to trying it out, each revision so far has made massive leaps forwards - really impressed!

WD

[wonkydog]

Showing now... all sorted 👌

[JimBuzbee]

Regarding Frigate+ models - When should a detection be submitted to Frigate+ ? For example, if it is having no issues detecting cars, should car detections no longer be submitted? Or if it typically has no issues, but If get a false car detection, should that false one be submitted and no others? And if I really want to get good detections for people seen in my driveway, should I keep submitting detections even when it seems to be doing a good job?

[NickM-27]

it doesn't help much to submit images that are already trained well (assuming all objects are trained well).

So, for example, if you have an image with a raccoon in it and you need to train on those - it is fine if there are cars in the image that already score well.

But, I believe it is not a good use of time to submit images that only contain objects that score well. Best to focus on cases where objects score with low confidence (maybe further away, or while partially occluded , for example).

[naitkris]

I submit images where the % score is less than 80% correct however it is labelled correctly - the idea being that by submitting it will help in future to increase the % score to be 80% or higher. Not sure if what I am doing is the best way but that is what I am doing and it does seem to improve things for my own fine tuned model.

[rpm5099]

I have a model that I trained about a month ago so I need to update, but I'm never sure what objects my currently loaded model supports. I've tried adding other animals in the past with prior models that were listed on frigate+ as being supported only to get an error on startup. What is the best way to enumerate what objects your current model supports so you don't have to play the guessing game in the config?

Similar question the the previous - I've been having relatively poor luck with fox and raccoon and it seems to have gotten worse with the last two F+ models I've trained - both within the past 6 months. I've got very good exemplar images of all of the objects I really care about, do I need to submit additional images when I request my latest model or will it retrain the latest base model with all of my existing images? Ideally I'd like to train the new base model on the currently existing images, and from what's been said my results should be better.

Great news about the new model, really looking forward to trying this out - please keep it up!

[NickM-27]

All supported labels are listed in the docs https://docs.frigate.video/plus/#available-label-types

if you get errors, then that suggests your config is likely not correct

do I need to submit additional images when I request my latest model or will it retrain the latest base model with all of my existing images? Ideally I'd like to train the new base model on the currently existing images, and from what's been said my results should be better.

this is answered in the docs, your images are used to fine tune directly from the base model every time, you can just submit a new request

[broyuken]

I just requested a model 2 days ago. Is there a way to get it retrained on .3 aside from requesting a new model again?

[baudneo]

In the email, it says that all currently requested training will automatically be updated to .3

If you have a recent model request, it has automatically been updated to use 2025.3.

[broyuken]

Yes but mine is not currently requested. It has been completed on .2. My question is, can I have it redone on .3 without using one of my 12 tunes?

[blakeblackshear]

It can't. The money on the GPU time has already been spent. There isn't a way to convert it to a different base model without retraining and using more GPU time.

[broyuken]

Ok no worries, I’ll submit a crap ton more images and request a trained .3 model next month :)

[borgqueenx]

Im curious, every time a new base model releases i read how it improves new labels. I might speak only for myself but human and car detection are my biggest use cases, have those been not improved in the last year releases? Has the peak of performance been reached for these?

[DrSpaldo]

Here are a couple from today that I am getting after a little training in the 2025.2 model.

which model type?

also, what are your thresholds set to for person? 77% seems pretty low

@NickM-27 Base 2025.2 with yolov9s on Nvidia 3060 using onnx detector. I am using min_score: 0.7 & threshold: 0.75 but only because of previous issues like @zonyl said that I was getting really low scores for actual people. As you can see though in the picture I posted the second one detected at 92%, so if I had it set at 80 or whatever, it still would have got it wrong...

[NickM-27]

sure, I am not suggesting it would be without error, but also having a low threshold like that you will have more false positives. I personally run with this, detects people on bicycles far away which is about as extreme as it gets on my cameras (this one was 86%).

objects:
  filters:
    amazon:
      min_score: 0.8
    bear:
      min_score: 0.5
      threshold: 0.8
    bicycle:
      min_score: 0.65
      threshold: 0.8
    bird:
      min_score: 0.5
      threshold: 0.8
      max_area: 80000
    car:
      min_score: 0.65
      threshold: 0.85
      min_area: 500
    cat:
      min_score: 0.65
      threshold: 0.85
      max_area: 25000
    dog:
      min_score: 0.6
      threshold: 0.85
    fedex:
      min_score: 0.8
    package:
      min_score: 0.55
      threshold: 0.8
    person:
      min_score: 0.6
      threshold: 0.85
    rabbit:
      min_score: 0.6
      threshold: 0.8
    raccoon:
      min_score: 0.5
      threshold: 0.7
    squirrel:
      min_score: 0.65
      threshold: 0.8
      max_area: 80000
    ups:
      min_score: 0.8
    usps:
      min_score: 0.8
    waste_bin:
      min_score: 0.65
      threshold: 0.85

[NickM-27]

to be clear, just sharing this for more info, seems interesting that there is this much variation in performance, I do have almost 8000 images annotated though

[DrSpaldo]

@NickM-27 thanks, I do not have any where near as many annotated. Especially for that side camera, as it doesn't get the traffic like my street facing ones do. I will however give masking those bushes and change the person threshold to what you have and see what difference it makes. Along with the new 2025.3 model once it processes

[zonyl]

For reference (Im competing with Nick ;) ) Thanks for sharing your thresholds! I have my min_score for people at .7 and threshold at .7 I guess I could raise threshold a bit based on median.

[mattenz]

Could we get "Police" added as a label? Would be kinda interesting to know when they're in the neighborhood....just because I'm nosey, not because I've got anything to be concerned about, of course

[zonyl]

I have firetrucks responding in my neighborhood occasionally and was thinking about "emergency vehicle" somehow. Not exactly sure if I would take any sort of action other than just curiosity.

[DavidValeri]

I'm looking forward to the .17 beta's becoming available to have a play with that :)

But, while the styling of every police vehicle is different, in most English speaking places I guess the common denominator is that they all have "Police" written somewhere, right? Would the model not be able to pick up on just that label?

I've lived enough places to know that it is all over the place. Police, Sheriff, State Trooper, Highway Patrol, and probably more. Colors too: black, gray, white, brown, champagne, etc.

[zonyl]

While I agree just identifying by the word police would be fruitless, there are likely less variations of police cars than people and clothing 😁. I was actually very impressed by CLIP search in Frigate identifying firetrucks very well

[hawkeye217]

There's an important difference between object detection models (which Frigate+ is) and CLIP models.

Frigate's object detection models are specialists: they are rigidly trained on a fixed, hand-labeled list of objects and require seeing thousands of examples of a specific fire truck design to recognize it reliably. Training for every global fire truck variation is infeasible for a detection model.

The Jina AI CLIP model used for semantic search is a powerful generalist. It was trained on hundreds of millions of image-text pairs, learning the concept of a "fire truck" by associating images with diverse descriptions like "red emergency vehicle" or "truck with a ladder." This gives it zero-shot recognition capability, allowing it to easily generalize and recognize a visually unique fire truck it has never explicitly seen before simply by matching the image's semantic meaning to the search term.

[NickM-27]

there are likely less variations of police cars than people and clothing 😁.

that's not really a valid comparison though, because Frigate does not detect a person based on their specific clothing, just like it does not detect a specific car based on what it looks like.

The delivery company and in this case suggested police detection would be entirely based on that characteristic

[Fswalle]

hi there folks,
i´ve got frigate+ and trained models for my coral, runs very well. for testing i got a yolo model, didn´t know the name, but it was too much for my system, so i went back to coral mobiledet model.
now i would train a couple of new pics, and the plus site shows me, that all my suggested pics are untrained and not verified.
do i have to verify them all again ? nearly 10000 pics ?
not really ...
thx for answers

best regards
franz

sry for my bad english, its´s school learned at 1981 ^^

[detroitgoose]

Yes, you have to re-verify all images when adding new labels.*

*If you don't need a new label for a specific camera, don't enable it for that camera and you won't have to re-verify that camera's images. So for me, since I'm not planning on using any of the new labels, there will be no re-verification for me to do.

It's a bit of a pain when it happens, but it can be made easier by verifying using the SPACE bar. Honestly all of the hotkeys really help speed up annotation (if you aren't already using them.)

[NickM-27]

To be clear, you do not HAVE to re-verify. The training will still use all of your images if you request a new model. But if you want accuracy for the new labels you enabled then you should, at your own pace, go back through the previous images to ensure those labels are annotated.

[Fswalle]

ok, that's perfect for me. the most objects are good trained.
there are only a few pics of fox on 2 of 7 cams, and that's no problem for me, because this cute one has just discovered my property as a hunting ground and is now coming several times at night. New model is requested and i'm sure, that it'll be the next step to Victory 😉
again thx a lot
best regards

Franz

[blakeblackshear]

You can also see some of my comments on previous related discussions:
#19053 (comment)
#18110 (reply in thread)
#18305 (reply in thread)

[Fswalle]

thx @blakeblackshear,
i'm fine with it now, was just a little bit fearful...
😉

best regards
Franz

[matt-garman]

I have also added many new candidate labels requested over the past few months. These include: ... coyote ...

I have about 6k submitted images, many of them with dogs (most being of my own dog). I do have a a handful of submitted images with actual coyotes. They were detected as "dog", and I submitted to Frigate+ with "Yes this is a dog" and verified as such - which is true if the word dog is used generically to mean canine. I suspect I'm not the only one to do this.

The fundamental question is, should I go back and relabel these coyotes as such?

More generally, I'm curious about any model's ability to consistently distinguish coyotes and pet dogs. My own dog still sometimes gets mislabeled as a cat. I can imagine some pet breeds like German Shepard or Malamute getting loose, and, depending on picture quality and contextual hints, a lot of humans would have trouble distinguishing them from a coyote (or wolf or jackal).

A more general question - as the number of images submitted to Frigate+ grows, when we opt-in for a new label, re-verifying images becomes increasingly tedious. For example, I recently switched from mobiledet to yolov9, and I got my first skunk detection. That was with about 1500 images submitted for that camera, so it was pretty tedious to go through all those and re-verify all of them for the single skunk. On the other hand, the second time through I did catch some stuff I originally missed. (Please don't take this as a complaint - just looking for a discussion on how to tackle this problem. In other words, I'd like to go back and fix the coyote labels, but it's a lot of time to sort through all the images; and, personally speaking, I'm content to have all canines detected as dog.)

(Edit: fix typo)

[zonyl]

I have almost the same situation. We have cats, dogs, foxes, racoons, coyotes, possums, otters, captured almost every day.

From a year ago to today, I am extremely impressed with its growth in ability to differentiate between similar animals (9K images myself). I was getting cat / dog mixups routinely until this last 2025.3 base model update which has made it 60/40 to I would say 80/20. Having said that though I was always very careful to never mislabel a coyote as a dog or possum as a racoon. I would just submit negatives and no false positives as true.

Just had a few foxes run by last night and it picked up those correctly which was a recent bump in quality as well (usually they would appear as dogs / deer).

My suggestion for adding labels would be to Re-run the Frigate+ bigger model on all images when a label is added ( candidate and 'trained') and pull out images for verification based on the following:

• New Objects detected.
• Different Object detected.
• Confidence change within a threshold.

• Have a buried (danger / warning screen) auto approve everything else

I think there is a chicken / egg scenario though as from what I understand the candidate labels are not trained yet when published.

[NickM-27]

To be clear the suggestions model is already run on all images for a camera when you add a new label (non candidate) to it

[NickM-27]

As far as I understand more improvements and other features are coming to this workflow in the future

[zonyl]

To be clear the suggestions model is already run on all images for a camera when you add a new label (non candidate) to it

Hmm. I got a total of 5 images in suggestions out of 8 or so cameras I added both new and existing labels to this go around. Maybe the annotations were just that good already. Definitely a candidate label issue then.

[blakeblackshear]

I'm curious about any model's ability to consistently distinguish coyotes and pet dogs

Me too. I think the answer is different for the base model than it is for user models. The base model struggles to differentiate between cats and dogs. I believe this is because there is so much variation in how cats and dogs look. There is a ton of variety in dog breeds. However, when a user fine tunes the models with their own images, it becomes biased towards the dogs and cats that are seen on their specific cameras and it is usually able to better differentiate. I suspect the same may be true for coyotes, mostly because coyotes are probably always seen at night on cameras and generally look similar to one another. I won't know until there is enough training data to try.

I'm content to have all canines detected as dog

Then I wouldn't go back and change them, as you won't get any benefit. Just wait until it's supported and see if the suggestions find any in the older images.

[jeeftor]

Is there a bug in the GUI?

What does CURRENT represent?

[zonyl]

AFAIK, it is the count of images with that label across all cameras, currently.

[blakeblackshear]

The "Current" count matches the "Images Verified" column. "Last Model" shows you how many you had when you requested your last model. The only time you would want to submit a request with that few new images would be with a new base model like this.

[notmaroun]

How recent is recent RE: model upgrades for recent requests? I submitted a request on 17/10 - but it seems to have stayed on 2025.2.

[AML225]

#20552 (comment)

[blakeblackshear]

Models that finished processing on 2025.2 cant be converted. Only models that were still in the processing queue were able to be upgraded. There is no way to swap out the base model after training completes. It has to be retrained.

[sapanbehar]

how are you guys submitting false positives?

[NickM-27]

https://docs.frigate.video/plus/annotating#false-positive-labels

[traveler2100]

I am running the mobiledet model on two TPUs with 10 4k cameras. I am considering upgrading to a hailo8 and yolo9 models. If I understand correctly, new base models include the fine tuning images previously submitted. Will the YOLO9 model include my previous images? Since I have to specify the model when I submit images for training, I suspect they are not included in the YOLO9 models and I'll need to start over. You guys with experience running Hailo8 and YOLO9, do you recommend the change from Coral TPUs and the mobiledet model?

[traveler2100]

Thanks, I was hoping to be able to go to 640 detection with the Hailo. My detection is good, running about 6 to 7ms, but I think 640 will give me more range, but sounds like it's a struggle even on a Hailo. Maybe we'll get a medium in the future.

Given my speed and good detection, I am not sure a hailo would be a big step up if I still run a 320s model. I am sure the Yolo9 model is better but there is not a lot of room in my implementation to improve. i'm running in docker on an ASUS core 5 nuc with Home Assistant in a KVM VM on the same machine and it runs pretty well.

[NickM-27]

but I think 640 will give me more range

No, 640 is worse for far away objects than 320. Which makes sense because 320 is more zoomed in.

[traveler2100]

That makes sense, but I hoped that having more pixels in the image would make it more recognizable. If it is indeed worse at longer range then 640 must have very little advantage over 320 since 320 does pretty well at close range and 640 uses a lot of resources. Of course with a hailo I may be able to up the detection resolution from 720P to 1080P which should sharpen up the detection box.

[hawkeye217]

You may be misunderstanding how detection works. Your camera's resolution determines the raw detail available in the image (see the docs on choosing a detect resolution), affecting overall quality and the processing load for initial decoding and motion detection. Conversely, the model's inference input dimensions (eg, 640×640 or 320x320) is the fixed size that regions of interest are resized to before the object detection model processes them.

[traveler2100]

Thanks Hawkeye but I recognize the difference in image substreams used for detection and the inference resize box. I also think 640 will do better recognizing small objects at a distance and I think that starting with a higher resolution image may improve the detection for facial recognition, license plates and small objects at a distance, especially at low light....but it is not something I have tried so just trying to see what you guys have experienced.

[ampersandru]

for new candidate labels, I should not add them in frigate config, correct? as they are not in actual Frigate yet?

[NickM-27]

It won't hurt anything, but yes they won't be usable

[choppedpork]

has anyone else experienced a relatively drastic increase of false positives for 'person' on 2025.3? i'm currently getting 10-20 alerts for a person daily for each of my cameras vs pretty much zero before the new model (other than real positives).

fwiw this is the first time i have not tried to use the newly released base model but instead processed some more of my image backlog before requesting a fine-tuned model. this added about ~450, totalling at 2.2k now - most of which was adding real positives for birds, foxes, squirrels and cats for which i've had lots of false positives before and so trying to balance things out. my previous attempts at using the base model all resulted in lots of false positives - might be worth noting that this is almost always "extreme" false positives, not objects of label 'person' being misidentified as 'dog' (those did happen to me too but they are very rare) but just random bits of foliage, branches, lens flares, potted plants etc. this is the first time i've had this much false positives on a fine tuned model (possibly for other labels too - i mostly notice 'person' because for these i get home assistant notifications).

i'm still on mobiledet/edgetpu because yolo on my gpu results in 10x+ more energy consumption (~2.5W for the coral vs additional ~30W for the gpu when using yolov9t onnx 320) while inference is twice slower (~8ms for the coral, ~16ms for the gpu) 😅 detection is at 5fps, 1280x720, motion masks in place as much as practical.

[naitkris]

I have the exact same problem now since updating - at night so many spider webs near the cameras are detected as a person.

[choppedpork]

fwiw i've always had a problem with spiderwebs and i try to clean the cameras as often as feasible, depending on their position they can entirely obscure important areas at night due to the IR light being so close and thus the webs getting so bright. they are also slightly seasonal (at least where i am) so i'm usually not surprised by these alerting me sometimes. the number of notifications for these is definitely quite a bit higher now though.

so far in the past hour i got two false person positives since switching to the 2025.3 base model, won't really be able to properly compare until at least 24hrs have passed (most false positive detections tend to be clustered around a particular time or all night long presumably due to light direction and level changes.

once the base 2025.3 model has had 24hrs i will try reverting back to a finetuned 2025.2 as well, to try and rule out it's a seasonal issue rather than something about the 2025.3 model seeing people in more places.

[choppedpork]

well, the one thing i'm sure glad about is that the base model is performing worse (~250 false positives in the last 24hrs from one camera, ~30 from another) so it's not the extra 450ish images i've added to the set that have caused the massive increase in false positives.

i'm going to switch back to 2025.2 now as i get better results with it in my environment (and none of the new labels are useful in my case) but i wonder if i should add a bunch of the false positives i've gotten with 2025.3 and try training a new fine tuned version? i usually try to batch new images up once a month but given 2025.2 won't be available for fine tuning anymore maybe using up my training credits on getting 2025.3 work better might be wortwhile... 🤔

[DRazumovskiy]

Have the same issue. With 2025.3 the score of false positives for people have jumped up to above 90% and filters do not work anymore. I have around 900 true positive and 700 false positive people images uploaded to Frigate+. I've also noticed that objects that I specifically trained as false positive started to pop-up as tracked with high score after I switched to 2025.3.
Few examples classified as people:

top score 91%

top score 82%

top score 81%

top score 91%

this one is especially frustrating, because I have at least 10 similar images used as false negative in the training dataset. Still got a top score of 90%

[NickM-27]

I have around 900 true positive and 700 false positive people images uploaded to Frigate+

It is worth pointing out that this far exceeds the recommended ratio, and not all of your false positives may be used in training each time. https://docs.frigate.video/plus/#improving-your-model

[bagobones]

2025.3 (Base Model) yolov9t (320x320) waiting for next month when my training subscription resets to build a new fine tuned model.

All bears and racoons are VERY confidently my cat or a stray cat. I do have racoons and bears in my area so the label is relevant to me.

bear:

racoon: ( I have adjusted my threshold over 85% now for racoon but still that will probably not cover it)

[ZhaiSoul]

Should we consider adding "mobile phone" as a candidate label? Sometimes I often forget where I've left my phone in the yard, and I believe this label would be very meaningful.
For example, in certain special situations, it might be necessary to determine whether someone is using a mobile phone?

[bagobones]

Mobile phones are for the most part black rectangles these days I am not sure that would be an easy thing to correctly mark label at a distance

[ninze]

I understand the reasoning behind re-verifying images when new labels are added to a camera. However, reviewing thousands of images simply to confirm that a particular object does not appear in them feels excessive.

The main issue is that partially verified images (where some labels are confirmed and others are not) are grouped together with completely unverified images. As a result, when I add a new label to a camera that already has over a thousand images, I end up with a large backlog of images requiring review. Then, if I encounter a false positive and submit it, that image gets mixed into the same backlog as those partially verified images, making it difficult to manage and track progress.

A separate tab for partially verified pics would probably make life easier for a lot of people.

[blakeblackshear]

That should be easy enough to add. In the meantime, you can quickly filter down to the ones that have never been verified by using the label filter to choose something like person that has always been enabled for the camera.

However, reviewing thousands of images simply to confirm that a particular object does not appear in them feels excessive.

I would suggest that you don't go back and label thousands of older images. Just review the suggestions on older images. I would only do this if it was critically important for me to have high accuracy on the new labels. For candidate labels, just wait until they are supported and suggestions are available.

[yousaf465]

No donkey label included?

[ninze]

@zonyl You can probably reduce false positives caused by spiderwebs by setting min/max area of the objects.

I also have this buddy hanging around the camera all the time, causing endless motion detection during the night, although I managed to decrease false positives to almost zero. Eventually I will probably move the camera a little bit further away from the roof edge just to save unnecessary processing time.

[zonyl]

@zonyl You can probably reduce false positives caused by spiderwebs by setting min/max area of the objects.

I also have this buddy hanging around the camera all the time, causing endless motion detection during the night, although I managed to decrease false positives to almost zero. Eventually I will probably move the camera a little bit further away from the roof edge just to save unnecessary processing time.

A good suggestion however as you can see in my images the actual detection is approximately the size of a typical person in frame. I have chemical protection on each camera and unfortunately am seeing spiders and webs father away. We have Joro spider infestation problems here and their webs are 20-30ft+

[ZhaiSoul]

When the bounding box of a recognized object is significantly larger than the object itself, or only part of the object is detected (e.g., when the full person is in the frame but only the upper or lower body is recognized), should this be considered a false positive? Also, in cases where the object is occluded or not fully visible in the frame, should it be labeled as "difficult"?

[bagobones]

It is supposed to predict the shape and size of the object even if obstructed. When labeling you are supposed to guess the full area of the object even if obstructed.. I have gotten this wrong in the past but this is covered in the frigate+ documentation.

[ZhaiSoul]

Thank you for pointing that out. However, I'm referring to cases where the object is near the edge of the frame, making it impossible to annotate completely. For example, when a person appears at the very bottom of the frame and only the upper body is visible.

[NickM-27]

If the box is on the object at all you can label it as true and then correct the box. False positives are only needed for cases where it is the wrong object type / not an object at all.

As far as difficult goes, it is generally subjective based on if you think it is not reasonable to expect the object to be detected with high confidence. Currently difficult does not do anything.

[ZhaiSoul]

If the box is on the object at all you can label it as true and then correct the box. False positives are only needed for cases where it is the wrong object type / not an object at all.

Thank you very much for your reply. Would it be possible to consider adding this content to the documentation and the UI? It could help better avoid misunderstandings.

[bagobones]

FYI this model and the suggestion model on the frigate+ site REALLY strongly prefer to label my cat as a dog now.

[blakeblackshear]

That's interesting. The suggestion model was not updated for 2025.3. It is still using 2025.2.