Frigate+ Model Launch

[blakeblackshear]

I am excited to announce that Frigate+ Models are ready to be rolled out. I can't believe it has been over 2 years since the initial discussion about the idea. Thanks to everyone for your patience and encouragement over that time.

edit: Subscriptions are now generally available
I will be inviting users to sign up for a subscription incrementally to ensure I can address any issues and not overwhelm the system. If you would like to be invited, please send me an email at [email protected] from the email address associated with your Frigate+ account.

After much consideration, I decided on annual subscriptions as a pricing model. You can think of the annual subscription as the initial buy in to cover the costs of Frigate+ and the development and maintenance of the base model. However, I also implemented a credit based system for training that I think everyone will be happy with.

• With your annual subscription of $50, you will get 12 training credits with the idea that you can fine tune a model monthly
• I will also be adding the ability to purchase additional packs of credits for $5 each
• It will take one credit to train a fine tuned model
• Models trained are yours to keep and use forever (under the terms and conditions) since you paid for the resources required to train it

Before subscribing, please read through the documentation on how the models work so you are clear on what you are purchasing.

[T3chArmy]

Very excited for this launch! Can't wait to continue to see this project grow!

So, I have a general follow up question, I feel like I already know the answer to this but would like a definitive answer.

Since the current Frigate+ Model only has a small subset of labels, if I include a label from the default model are those still detectable? or does the Frigate+ model completely override the default model?

[T3chArmy]

Thanks for following up! That's what I was thinking, but just wanted to make sure!

[dopeytree]

Is there a way to merge frigate normal & frigate+?
For example Whosatmyfeeder sends the initial detection to another container to do further analysis using another model and then returns to frigate with a sublabel.

PS I will be signing up and paying but would like to keep some other labels.

[NickM-27]

No, only one model can be used at a time. To use whosatmyfeeder along with a frigate+ model you would need to wait until frigate+ supports the bird label

[dopeytree]

Hmm my solution could be I use 2 instances of frigate. Perhaps run 0.13 beta with frigate+ custom model & then keep frigate 0.12 for sole bird detection on the google model. I have a camera dedicated to it.

[NickM-27]

yeah, if you have a camera dedicated to it that would not be a bad approach, although keep in mind you can only use a coral for a single process so a single coral could not be used for both

[shinyzard123]

What if I don't feel comfortable uploading my own images? Is there any possibility of just paying a one time fee to use the base model? I know the experience won't be as good but I'm sure it would be better than the current one

[blakeblackshear]

Not currently. You could upload and verify just 10 images from your cameras with either a false positive or images that don't have any person or car objects that you are concerned about sharing.

[shinyzard123]

Ok, that's fair, thanks for the answer 👍

[T3chArmy]

Another follow up question,

I've requested a Frigate+ model to be generated, will I receive an email once it has completed?

[NickM-27]

Yes, looks like it

[T3chArmy]

Excellent. Thanks!

[NathanVanCleave]

When annotating uploaded images, how does selecting the "Difficult?" checkbox affect it? Does it simply increase/decrease how it is weighted? Just want to make sure I'm using it correctly. If an object is more than half obscured (I still select the whole object) or relatively small I've been marking it as difficult.

[NathanVanCleave]

Along these lines, how should we go about uploading images of missed objects for annotating that didn't create an event? I know I can upload my own images, but I'm unsure if screenshots are appropriate or if I need to somehow export them from Frigate for standardization. I've been uploading old snapshots Frigate saved from months ago, but those are only of tagged events. As long as the screenshot resolution isn't worse than the saved/uploaded snapshot I'm guessing it's okay?

[NickM-27]

I did this some by downloading the clip on my mac and using the save frame as png option and then uploaded a few of those spaced seconds apart.

Also, technically in 0.13 the api exists to pull a snapshot from a specific time in the recordings

frigate/frigate/http.py

Lines 1238 to 1239 in 0858859

	@bp.route("/<camera_name>/recordings/<frame_time>/snapshot.png")
	def get_snapshot_from_recording(camera_name: str, frame_time: str):

This is something that will be supported directly in the ui in a future version of frigate.

[NathanVanCleave]

Okay, saving the frame as a PNG is something I can do. It sounds like you're at least recommending I do that or use the API instead of a screenshot then? I'm only pressing the question because it's the easiest option in the moment if I happen to catch it missing something, but it's not worth the time saved it's it's not an equally valid/confirmed option as the former.

Happy to hear this feature for the UI is on the roadmap.

[NickM-27]

Personally I don't like screenshot because it is difficult to get the edges right and the resolution will be from your screen not from the original source, plus any other compression that may happen. I don't know that it would actually matter for the training, but to me it's not difficult to use a unix timestamp and pull multiple snapshots via the api

[NathanVanCleave]

That's exactly what I was thinking about the screenshot being different. This is good enough confirmation for me not to do it this way.

I don't know how to use the api, but I'm fine with downloading the clip and saving the frame. Thanks for the replies.

[agrieco]

Probably too early for this, but it would be nice to know how much the underlying training data has changed. For instance- counters on how many images have been uploaded and used to train the new model since a user last built their own model. Maybe even by label?

In the end, I'm trying to better understand when the community aspect of the Frigate+ subscription would warrant a new model build.

[blakeblackshear]

New base models aren't automatic. I will be hand crafting those for the foreseeable future. When the next version is available, I will be making some updates in the UI and notifying users. I can show how many of your verified images were used for each model in the next update.

[mabed-fr]

Great !

Has this functionality been addressed? In which version will this be deployed? I will wait for this version to subscribe.

[mabed-fr]

Is it possible to publish the number of base images and other statistics on the base model of frigate+?

[blakeblackshear]

I am testing a new base model now and it does appear to be an improvement. Just making some changes in the backed to support the roll out. After the first few updates the process will get smoother, and I will be able to update more frequently.

[borgqueenx]

how much percentage would you say it improved? (ofcourse a rough estimate) got my coral today, want to spend my first credit when the new base model is out and start training.

[Fervere]

Ive subscribed. But dont want to request a model.. because I have some cameras with some pictures on them and they dont show up.
And on the main camera I want to test this out on, I had an old camera, and that also have some pictures on that Ive verified earlier.

Is it ok to mix cameras, they stand at exactly the same place but the angle might have changed a bit?
How to remove images? please add, if im blind.

If i request now, is it possible to start from scratch later?
What if I move my camera a few meters should I add a new camera and upload new pictures?
Is it possible to remove some images from the model later?

In 1 of my images now ive marked 1 person, and in the background there is a bike with a person on it. And I cant select bike anymore, what to do?

Why not allow us to add annotations on things we think is relevant so these pictures can be used later?
Id like to add my robot mower, should be simple.

[blakeblackshear]

It's ok to mix images from cameras even if the angle has changed a bit. Your model is a tuned layer from your images on top of the base model trained on a collective set of images, so your images aren't the only training data.

There isn't currently a way to remove images yet, but it's a common request, and I am working on it. Images removed later won't be used in models trained after they are removed.

There is not currently a label type for bike in the Frigate+ model set, but it is in the list for the expanded set of labels. Frigate+ models only detect a specific set of labels currently.

[NeoDuder]

Can I assume a single model covers multiple cameras?

[NickM-27]

a single model is trained on all images that you have submitted, labeled, and verified so that would include multiple cameras assuming you have submitted, labeled, and verified images from multiple cameras.

[NathanVanCleave]

What's the purpose of tagging the images by camera then? I guess I assumed the models would be somewhat specific to cameras, though that would require multiple credits I suppose. Could it be possible/beneficial to request models per camera? I've been aiming for the recommended 100+ images per camera before I request any. Maybe it's more about focussing on the hardest images (e.g., backyard with a far field of view) and not worrying about easier ones as much (e.g., doorbell where subjects are usually very close).

[NickM-27]

I'm not sure what you mean by tagging the images by camera.

But yeah I personally wouldn't expect each camera to have the same number of images as you start to use the model, this is my distribution for example

[NathanVanCleave]

That screenshot is exactly what I mean. Thought it had some special significance. Good to know.

Your distribution is exactly what I'm aiming for.

[tpjanssen]

If a request a Frigate+ plus model, will it be based on the standard model that's been trained additionally with my submitted images? The default model is pretty bad in making a difference between a cat and a dog. Will the downloaded model only become better with my own submitted images, or is it trained with all the images submitted by other users?

[tpjanssen]

Thanks for the reply.

I haven't used the default model at this house / with these camera since I moved a few months ago, but at my old house my dog was labeled as person, cat, and bird by the default model at least a few times a day.

This is what I experience as well. Frigate also identifies people as cats or dogs from time to time.

I can imagine that the base model evolves and gets better the more it's trained. What are the demands to use the extra trained model? Only the submission of 10 images? Or do you I also have to subscribe (which I will certainly do),

[NickM-27]

I think that was answered here but I would definitely personally suggest submitting images as that will make your model better than the frigate+ base model for your cameras.

[tpjanssen]

Thank you Nick. Spent some time on labelling and veryfing all my events which were identified wrong. Did some of these about a year ago, but back then it didn't make much sense. Good to hear that nowadays it does.

[pyrodex]

I submitted a few today of my wife and kid bending over triggering a dog, it was too funny.

[pyrodex]

Here is a weird example showing the steam from my tankless hot water heater this morning being classified as a dog.

[timmyag]

I might be slightly off topic,
Would having sperate models for cameras in black a while/IR mode vs colour worth considering?
I mention it here as mine produce far more confused results in BW mode and im assuming that partly as the images arnt as good but also the training data is mostly in colour. Seems worth it to not train a model for use in the day on BW images and train a model for night time with as many BW images as we can.

[NickM-27]

I have had no problems getting great day and night detection with my cameras and the frigate+ model. The color of the objects may be different but the shapes are the same.

The important part is just submitting examples that are from all different times of the day so the model is tuned on your cameras at these different times.

Here is an example of a cat detected at night

[kirsch33]

i wish it were possible to delete images from Frigate+. i uploaded a bunch of just absolute crap photos that I think are probably confusing the 1st model that i created, and would like to be able to prune these out

[bagobones]

I am about to upgrade a bunch of my cameras and being able to delete out whole camera sets and start replacing them with higher quality images would be very useful.

[ccutrer]

As long as your images are labeled correctly, I wouldn't delete them just because they are lower resolution. Remember that some number of images are taken from all contributors to train the base model, and you'll be helping out everyone.

[bagobones]

As long as your images are labeled correctly, I wouldn't delete them just because they are lower resolution. Remember that some number of images are taken from all contributors to train the base model, and you'll be helping out everyone.

I disagree at least from my experience with training faces with other models. Crap cameras will introduce a lot of blocking / mpeg artifacts to moving snapshots.

I found that if I trained off of certain camera my faces would more often match many people than individuals. My older cameras also tend to ghost more when in IR mode.

Doing some other research it seems recommended to train from a higher quality source when possible.

[NathanVanCleave]

How important is including false positives in annotations? I just got my model last night, trained on ~400 images for four outdoor cameras, and am getting a ton of false positives so far I wasn't getting before. Since the beta is pretty new I only had false positives included in <5% of my images.

If false positives are important for training, I'll hold off tweaking my masks to continue allowing them and continue uploading more images for annotation before requesting another model.

[tpjanssen]

After reading all this I think it is better to switch to the base model, and do some consistent reports next coming months. I'll try to have a good mix of true and false positives, along with a good spread of day/night reports.

@blakeblackshear is it possible to add the boxes of all detected objects to the submit function? Maybe you should not enable them by default on the website, but this will make the verification process a lot easier!

[NickM-27]

You would have to store the location of all of the objects at the moment that each snapshot is taken, that would be a lot of data to store and to keep track of. I believe Blake has already said there are plans to improve this process but there are better ways

[blakeblackshear]

is it possible to add the boxes of all detected objects to the submit function?

It's not possible. All that data is only available in real time. By the time there is an event with a submit button, all that information is long gone. My plan is to scan the image for objects again in the submit to frigate+ dialog. Eventually, the entire annotation tool will be directly in Frigate.

[blakeblackshear]

@tpjanssen

I also have some privacy concerns about that. @blakeblackshear what can you tell about this?

The privacy policy and terms should answer any questions you have. If not, let me know what questions you have.

[didi3r]

Hey I just wanted to give an update on how my detection is doing. After 2 weeks I’m by my 3rd model with a total of 1700 images and the detection has gotten way better to the point that it now detects people I even struggle to see at first glance. Like this one where it detected a person in the dark:

Or this other one where it detected a person far far away in the corner of the block:

And they are not isolated cases the model is now constantly detecting at that level of precision.

I still get some false positives during the day, but definitely way less than the default COCO model. The approach I followed was to submit true positive detections with low scores or with incorrect bounding boxes and submit pretty much all false positive I got no matter if the images were almost identical. I would say that out of the 1700 images I have submitted, almost 500 are false positives.

Next steps for me would be to train the model on very specific areas of my interest so I’m going to be more cautious on which true positives to submit and reduce the
false positive submission a little bit.

[J-Pipe]

Since I record a high-res (main) stream 24/7, but my detect stream is low-res (substream), is there any concern if I manually upload some images from my main stream into frigate+ to tag some objects? i.e., should I only upload images that are the same resolution as the detect stream?

[NickM-27]

No, there is no concern. The model is trained entirely on a single size region (320x320) so resizing and cropping occurs anyway

[T3chArmy]

Now that I've had the new model running for a few days, is it possible to flag sub labels as incorrect? Have seen several things getting flagged as Amazon/UPS/FedEx/License Plates that are not.

[TyzzyT]

You can do it just like regular labels. Upload it to Frigate+ and annotate the image

[blakeblackshear]

You can't specifically submit a false positive for the sublabels because they don't actually get their own detection region. Submitting the primary label and verifying the image will help.

[tpjanssen]

While running new the plus model I've noticed that my interference speed significantly dropped (from around 8-9 ms to 6-7 ms). Why is this? Because the model is not trained with as much labels as the standard model? Or the labels itself are not as "heavy" (trained with less images) as the standard model?

The new model has the tendency to detect perons walking on my driveway based on the lower half of the body. Most of the time I see the bounding box drawn around the legs and feet only, with 90+% scores. As a consequence, this image is also used in the event list. What your advice to do with these cases? A couple of scenario's I can come up with:

1. Leave as is...
2. Submit, confirm it is a person, and draw an extra annation box around the complete body
3. Submit, confirm it is not a person, and draw a new annation box around the complete body

Is it possible to add the area and ratio information when you hover over the bounding box in the snapshot?

[tpjanssen]

Thanks guys. I will do as you suggested. What about my last question, is it possible to add area and ration information in the snapsnot view?

[NickM-27]

not sure I understand the last question, the snapshot is a static image. you can get area and ratio information from the recordings timeline overview. The area is already shown in the bounding box on snapshots (the number next to the % score)

[tpjanssen]

But these are not always the same. I see different bounding boxes for clips and snapshots. That having said, most information is already in the snapshot's the bounding box. It shows "person: 92% 53714". Maybe you can add the ratio there as well?

[jockelandelius]

Sorry for A revisiting an old post and B for stealing someone focus in this post but... "...get area and ratio information from the recordings timeline overview". I must be a total novice but I have never seen this and I am also unable to find it now that I see it does exists... Any pointers?

[NickM-27]

in frigate 0.13 you can click the icons at the top of the event recordings and it will show an overlay of where the object was detected. You can hover over this with your mouse to see information about it

[J-Pipe]

So if I want to rename objects in the labelmap, I assume there is a new map with different numbering than the COCO one? Not seeing where the labelmap is specifically documented.

[NickM-27]

it's not documented but you can find it inside the /model_cache/<id>.json file

{
  "labelMap": {
    "0": "amazon",
    "1": "car",
    "2": "cat",
    "3": "deer",
    "4": "dog",
    "5": "face",
    "6": "fedex",
    "7": "license_plate",
    "8": "package",
    "9": "person",
    "10": "ups"
  }
}

[J-Pipe]

Wonder why the example here renames 3/4/5 as an animal. Like the indexing is 1-based.

[NickM-27]

just an example, but those docs were also added a while ago and the indexes may have changed.

[blakeblackshear]

I believe that renaming the labels will currently break the true/false positive submissions to Frigate+. The image should still upload, but any labels that aren't listed in Frigate+ will fail. I will have to think about how to handle this more gracefully.

[tpjanssen]

@blakeblackshear what are your plans on new labels? The new base model seems to perform much better, but I have lots of birds detected as cats.

[mabed-fr]

Yes i want bird !

[NickM-27]

coming up soon https://github.com/orgs/frigate-nvr/projects/2/views/1

[tpjanssen]

Thanks Nick, missed that. Curious what the approach with respect to already submitted images will be...

[AdamPflug]

Cannot wait. This is my biggest issue as I have a lot of custom labels I'd like to use (chickens, racoons, rabbits, elk, tractor, etc). Even if they weren't detected/used in training yet I'd love to be labeling them so I could avoid a massive-relabeling effort later.

[derekcentrico]

I wouldn't mind foxes. Lol

…

On Tue, Apr 30, 2024, 3:22 PM tpjanssen ***@***.***> wrote: @blakeblackshear <https://github.com/blakeblackshear> what are your plans on new labels? The new base model seems to perform much better, but I have lots of birds detected as cats. — Reply to this email directly, view it on GitHub <#7932 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAOXG3TAD6U2TH3NXXWG65DY77VQLAVCNFSM6AAAAAA5E37EMCVHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM4TENZZGEYTM> . You are receiving this because you commented.Message ID: ***@***.*** com>

[Tyler-2]

I am one of the users using Frigate in a more natural environment. The new model seems slightly better in some ways and worse in others. I have a camera that fogs up at night and it results in (now with the new model) constant person and car detections. Trees in the breeze cause all sorts of hallucinations that didn't seem as common before. Stationary farm equipment / carts / wagons are often "car". A fairly stationary and uninteresting tree trunk is sometimes a " car". My chicken coop is full of "cats", "dogs", and "people". Confidences are high, >75%. I will just continue submitting false/true positives and negatives here and there (658 verified images so far) and hope it'll be useful for folks.

Chickens, foxes, raccoons would be great in my application, but I understand we are kinda weird.

[NickM-27]

to be clear 75% confidence is not that high. Recommended threshold is 85%

[derekcentrico]

I live in NoVA. Suburban life full of deer and foxes. So I appreciate the idea of more labels and later training.

…

On Wed, May 1, 2024, 12:11 PM Nicolas Mowen ***@***.***> wrote: to be clear 75% confidence is not that high. Recommended threshold is 85% — Reply to this email directly, view it on GitHub <#7932 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAOXG3Q2A32OZXIXZC5V3M3ZAEH4VAVCNFSM6AAAAAA5E37EMCVHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM4TEOBWG4YTG> . You are receiving this because you commented.Message ID: ***@***.*** com>

[nulkathos]

i think that labeling needs to be improved. sometimes it's difficult to label more than 2 objects that are near between them because the name/type of label hide part of the image.

[NickM-27]

just use the f key to hide other boxes. There are loads of keybindings

[NickM-27]

[nulkathos]

as always...thank you NIck :)

[f1d094]

@blakeblackshear What is the current wait time for a subscription?

[blakeblackshear]

I'm making some changes that I need to finish first. Planning to send all outstanding invites this weekend once I'm done.

[f1d094]

I haven't requested yet...wanted to see what the status was first. If I put in for a sub next week, how long would it be before I was up and running?

[blakeblackshear]

Normally I will send out an invite within the day. Sometimes within a few minutes.

[f1d094]

Awesome. I will send you $50 that day. Maybe within a few minutes.

;)

[dgikiller]

👋 Hey team, thank you for your hard work! 🚀
Quick question: I would like to share the account with my family to 10x the number of images correctly tagged for the first training session but it would be nice if all of us had a separate login.
Do you have any plans to create roles like image tagger on a shared account?

[mimainka]

Question towards the email asking for the invitation:
I'm using alias-addresses for practically anything, so my mail is like "[email protected]". But when I send you a mail it will come from something like "[email protected]". Would that be a problem for your invitation or can I just include the frigate+ email address associated with the account in the text part?

[blakeblackshear]

It doesn't need to come from the email you registered with. I just need to know what email you registered with.

[f1d094]

I've just started setting up Frigate+ and verifying my images but I cannot find any info for how to tag amazon, fedex, and ups. Should I just tag the word on a given truck? How about on the uniforms? Should I tag blue vest of amazon employees? Even if you cannot see the logo? The whole Amazon/FedEx/UPS truck?

Some guidance here would be welcome.

[NickM-27]

There is an example shown in the documentation https://docs.frigate.video/plus/improving_model#properly-labeling-images

[derekcentrico]

I've seen a demo where it is the whole vehicle and the logo tagged with the carrier's label. Nothing about the vests/clothing has been in the demo examples I've seen. I'd only assume you want to do the clothing SOLELY if you see the same logo.

[f1d094]

There is an example shown in the documentation https://docs.frigate.video/plus/improving_model#properly-labeling-images

@NickM-27 I missed this somehow. Thank you.

[f1d094]

Has anyone else had problems with long-running stationary object recordings with Frigate+? I deployed my trained model (1700+ images) and now have several long-running stationary object recordings.

Upon restart of the system, 3x stationary objects immediately are detected and start recording. Most notably my parked car in my garage, where there has been no motion today, and another car outside where there hasn't been recent motion.

The system in question has been running with the same config for several months. The only change was add my model and updating object lists on a couple cameras to add the objects unique to Frigate+.

Thoughts?

[NickM-27]

You might be confusing detecting objects and recording. Just because an event is ongoing doesn't mean it is continuously recording.

[f1d094]

@NickM-27 thanks for the clarification. That helps. Under Events -> Ongoing Events, these events seem like they are permanently there. When I click on the recordings the event is under a minute long, which speaks to your comment...so at least they aren't devouring disk space, but this begs the question: "Why are these events happening?"

When I select the cameras in question and look at the debug / motion boxes + bounding boxes, the objects are bounded correctly but there are no motion boxes.

Any ideas?

[NickM-27]

What you're describing is how it's designed to work. The new frigate UI displays this in a more intuitive way

[f1d094]

I also have discovered an unrelated issue. I already have a bunch of false positives for the new objects I'd like to start tuning on but I have no checkboxes to select to parse them out. The labels in the UI do not match the objects in my config. What I am seeing:

Objects in my config:

objects:
  track:
  - person
  - car
  - dog
  - cat
  - deer
  - ups
  - fedex
  - amazon
  - package

I have restarted Frigate multiple times, as well as cleared my cache and restarted my browser. Am I doing something wrong, is this a bug, or is this a known issue?

[ccutrer]

That just looks like the UI is allowing you to choose from objects that have actually been seen in the past, whereas the config is telling Frigate which objects should be tracked in the future. You're not continuing to have events for Mouse or Bicycle, right (regardless of if they are actually those objects, or just mapped incorrectly)?

[NickM-27]

Right, otherwise the UI wouldn't be usable to filter object types that still exist

[f1d094]

Ok, that makes sense. I never realized until today I actually have (false) events for mouse. Various categories come up empty for various cameras...I was under the impression that all categories showed up in the menu regardless of actual events. It wasn't so much that old categories were still there, but that the new ones were not.

The driver was that I was wanting to filter on "amazon" but now I realize that "amazon" only appears as a sub-property of "car". 100% of my amazon tags are not correct...I guess "fedex" "amazon" and "ups" objects not filterable then?

What is best practice right now for these three objects? Should I not use them until 0.14?

[NickM-27]

They will be filterable as sub labels

[pstephenwille]

I have a YoloV8 model trained that I'm experimenting with that recognizes playing cards and hands. I'm trying to get Events/Alarms fired when cards are detected. Would Frigate+ work for this? Could it use the Yolo model or does it need to be trained on the card images?

[mrbostad]

I would like to suggest adding a Annotation Label, i see there are Fedex, ups and Amazon, in Norway most packages and mail are delivered by "Posten" / "Bring" and some deliveries by DHL. Perhaps adding "Posten" for Norway and perhaps "DHL", or maybe make a generic label for package delivery services (containing UPS, Fedex, Posten, Postnord, DHL, Bring, Amazon etc") i.ex. "Transporters" or "Delivery Services"

Greetings from fresh Frigate+ user :)

[derekcentrico]

I label belgian postal logo as package and it is working pretty well.

You know doing "clever" things like this taints the model for everyone else, right?

No, because Blake will probably not use these images to train its base model.

Wait so the images of this "anonymous" naked guy I keep sending in as a package label won't be used?

[xavan83]

The logo is a red package :

[tpjanssen]

I label belgian postal logo as package and it is working pretty well.

You know doing "clever" things like this taints the model for everyone else, right?

No, because Blake will probably not use these images to train its base model.

Wait so the images of this "anonymous" naked guy I keep sending in as a package label won't be used?

😂

But @blakeblackshear can tell you more about this one... 😉

[Tyler-2]

Hmm. I figured our training inputs were used where confidence was high to update the base model, and that while more weight was put into your personal model for your personal trainining data, we are all still contributing to the base model.

Maybe thats wrong.

the logo is a package

Lol well can't argue with that I guess

[NickM-27]

User submitted images do not automatically go to the base model. Blake uses a lot of methods to choose the best images to go into the base model and also corrects any boxes, incorrect annotations, etc. This process is detailed in the post #9466

[lucius-the]

I actually decided paying for Frigate+ but then found the upgraded model has even less labels for animals then the default COCO one. I have a camera that regularly sees: wild boars and deers. I'd be also nice if I could keep detecting motorcycles.

Could the new label extension include custom labels - so that we can add our own ? I'm not sure if that's viable or not. I know too little about model training. If custom labels are not a possibility, could you please add "boar" and "motorcycle" labels.

[blakeblackshear]

Motorcycle is in the works already. The problem with purely unique custom labels per user is that you are unlikely to have enough examples on your own. I'm first focused on expanding the common set of labels for all users, and I will consider adding boar to that list. Custom labels per user is something I will look at later.

[lucius-the]

Thank you very much for your response and explanation. Looking forward to the next version and especially for a custom list. Do you have some "order of magnitude" number on how many pictures are needed for a successful custom label training ? Just curios...

[rsteckler]

Quick question about 14.0 and Frigate+ submissions.

I used to manually tag everything in the photo to make sure the model got the best inputs possible. In 14 the new submission window looks like this:

I clicked "That's a person", but I'm worried that the cars weren't tagged and I didn't have an option to add them.

Similarly, what would I push here?

That's a person, but the face is wrong. I don't want to tell the model that's NOT a person and inadvertently train that into the model, but I want to correct the face (and cars)

[NickM-27]

this works the same as in 0.13, you submit the image in Frigate and then you go to the Frigate+ webUI to finish annotating and then verify the image

[rsteckler]

ah gotcha. That makes sense. I guess I thought it was doing some of those steps behind the scenes now.