Frigate+ Base Model 2024.2 Update #13023
Replies: 16 comments 41 replies
-
|
When are you planning on implementing zooming? |
Beta Was this translation helpful? Give feedback.
-
|
I already have it working as a part of moving the annotator out into its own component library, so once I get that integrated back into the frontend. |
Beta Was this translation helpful? Give feedback.
-
|
Good stuff. Will we have the ability to scroll through images with the new frontend? |
Beta Was this translation helpful? Give feedback.
-
|
What I have in mind should cover the use cases I have heard so far and what's needed to go back and annotate with new labels. |
Beta Was this translation helpful? Give feedback.
-
|
This is great news, thanks Blake! What does "Fixed suggestions for objects consistently in the same locations" mean? I'm making assumptions, but wanted to clarify. |
Beta Was this translation helpful? Give feedback.
-
|
It means objects that are in the same location like a car that is often parked in the same spot can be a suggested box automatically |
Beta Was this translation helpful? Give feedback.
-
|
yes please, I often avoid submitting images from my front cameras due to all the cars visible and not wanting to annotate each one every time. |
Beta Was this translation helpful? Give feedback.
-
|
This is what I was assuming / hoping for! I'm very excited for this feature! |
Beta Was this translation helpful? Give feedback.
-
I have a DUO2 and live on a very straight section of street with houses all the way down as far as the camera can see.. I have SO many cars in view at all times of day.. |
Beta Was this translation helpful? Give feedback.
-
|
I mask my images before submitting to frigate+ with no adverse effects |
Beta Was this translation helpful? Give feedback.
-
|
Thanks for the update Blake - how often do you plan to do the base model updates going forward? Just trying to plan out how to use my trainings |
Beta Was this translation helpful? Give feedback.
-
|
I believe it says in the docs, but for a while now it has been said that there will be quarterly updates |
Beta Was this translation helpful? Give feedback.
-
|
From the Frigate + doc. . . INFO Is this likely to change anytime soon? I’d be willing to pay for a base model. |
Beta Was this translation helpful? Give feedback.
-
|
2024.1 has been great for me, I'd originally bought into Frigate+ in hope of reducing false positives due to cobwebs on the cameras and our shady looking BBQ cover in the corner of the garden. I can't even remember the last time I had a false positive now, excited to give 2024.2 a go! |
Beta Was this translation helpful? Give feedback.
-
|
I've the same problem (esp. at this time of year, worsening into Autumn!) and, given a few example frames below which are from an untrained Frigate mistakenly identifying "person" in each shot, do you think a trained model from 2024.1/.2 would improve things similarly well? |
Beta Was this translation helpful? Give feedback.
-
|
Turning off the IR on the camera and getting a seperate IR spot light might help more so the camera view isn't constantly blocked and needing a cleaning. |
Beta Was this translation helpful? Give feedback.
-
|
How to report false positives with the new UI? |
Beta Was this translation helpful? Give feedback.
-
|
you use the Frigate+ page in the sidebar |
Beta Was this translation helpful? Give feedback.
-
|
Ohh, great! Sorry I missed that. I did not have snapshots enabled so there was no images in the F+ sidebar page |
Beta Was this translation helpful? Give feedback.
-
|
Niiice! Thanks for all your effort! This is amazing timing! I've been procrastinating getting into Frigate+ stuff for half a year, and then this weekend something got into me and I've submitted 3000 verified images across 2 model trainings. The improvement is impressive. I imagine I may be a bit of an edge case because I put cameras on the ceiling of every room in my house and am trying to get Frigate to be a presence sensor of sorts. There are some unique challenges to it, like identifying people sleeping under blankets haha. It's already amazing! It's starting to be able to tell me apart from my dog even when I'm, say, hugging her on the sofa. |
Beta Was this translation helpful? Give feedback.
-
|
What about OpenVino support? |
Beta Was this translation helpful? Give feedback.
-
|
just tried openvino frigate+ model and it is amazing compared with the generic model. though it uses yolonas. working fine with my i7 7700T. |
Beta Was this translation helpful? Give feedback.
-
|
@blakeblackshear is an openvino model generated nowadays when you request for a new model? |
Beta Was this translation helpful? Give feedback.
-
|
when you request a model you specify if you want mobiledet or yolonas format |
Beta Was this translation helpful? Give feedback.
-
|
OK. This is added recently? |
Beta Was this translation helpful? Give feedback.
-
|
this past weekend I believe |
Beta Was this translation helpful? Give feedback.
-
|
From Frigate+ documentation:
A motorcycle is not available yet. Does it make sense to label a motorcycle as a car or a person to be able to detect them by a custom model or it is necessary to wait for future release with more object types? |
Beta Was this translation helpful? Give feedback.
-
|
I always label the person on top of the motorcycle (keeping the selection box around the rider only, not the whole vehicle). It's gotten nearly perfect at detecting that on my model. When we have motorcycle labels I'll probably select only the vehicle, so that it also detects parked ones. I'm sure that's not the only way to go about, but it makes sense in my mind. |
Beta Was this translation helpful? Give feedback.
-
|
Just figured I'd give some feedback... Night one with 2024.2, I've had 0 false positives for the first time ever. It is a little less sensitive to the true positives as well though, but the misses are generally what I would consider "hard" (far from being in the primary area of the frame, in the background in a distance). |
Beta Was this translation helpful? Give feedback.
-
|
My experience is very similar! Noticeable decrease in false positives. Lost some of the most impressive hard to get true positives, but they are edge case enough for me to think it was well worth it! |
Beta Was this translation helpful? Give feedback.
-
|
Really excited about the AI suggested label feature. Great work! Would also love to see the interface within the self-hosted portion of frigate automatically advance to the next picture when confirming/rejecting object detections, as well as keyboard shortcuts! |
Beta Was this translation helpful? Give feedback.
-
|
Can you please add this new to the frigate+ website? It makes it hard to hunt these upgrades. |
Beta Was this translation helpful? Give feedback.
-
|
How are things progressing on this front? Will any of the prerequisites for additional labels be available in the next point release? |
Beta Was this translation helpful? Give feedback.
-
|
I am almost done making the backend changes to track verification of individual labels on each image instead of a single verified flag on the image. There will be some small related UI changes when that goes live. There will no longer be a simple "verified images" count since verified is set per label. Once that is done, I need to update images view and annotator so you can review a set of images again for a new/existing label. Then I can add a lot of new label options for annotating. At that point, I can start incorporating new labels into the base training set so they are supported in the model. |
Beta Was this translation helpful? Give feedback.
-
|
Hello, Regarding the new model (2024.2): I've already re-train the model a second time with additional samples that were a missed true positive, but with no improvement. |
Beta Was this translation helpful? Give feedback.
-
|
Do you know roughly when the zooming feature will be available? I would like to train a batch of new images, but I'm holding out for the zoom feature first. |
Beta Was this translation helpful? Give feedback.
-
|
Here is the roadmap: https://github.com/orgs/frigate-nvr/projects/2
I have it listed on the annotator improvements already. I'm not sure how necessary it will be after suggested labels are added. I definitely don't want the images masked out like this directly in the image data. I wouldn't want to pull them into the base model, so I will need to think through the right approach.
Yes. I have already laid the groundwork for this so its possible to go back through images when you add a new label to a camera. If you add a new label to a camera right now, then all that camera's images will lose their green check.
I still need to add the ability to purchase more in 5 packs. There haven't been that many users who have run out, so I have prioritized other things. |
Beta Was this translation helpful? Give feedback.
-
|
Do you have an ETA on openvino support being finished? I am looking to setup another system at a different location and would prefer not to buy a coral if it's not needed |
Beta Was this translation helpful? Give feedback.
-
|
I would like the option to submit only a cropped object sometimes.. I have indoor and outdoor cameras and many times I don't submit due to personally identifiable info or family in the photos, but maybe I would like to submit a bunch of weak or false dog / cat detections. |
Beta Was this translation helpful? Give feedback.
-
|
@shinyzard123 in the email, blake told me around 4-6 weeks. should be the same time the next base model version release. |
Beta Was this translation helpful? Give feedback.
-
Suggested labels should help. This gives you an idea of what I'm working with regarding cars.
It would be good to have the option to scroll through images at any time to check labels and bounding boxes.
I'd be happy to purchase a pack of 5. |
Beta Was this translation helpful? Give feedback.
-
|
I noticed you are adding new object labels for the 2024.3 release. Not sure how to request for an object label to be added, but I'd like to add skunk to the list. Thanks. |
Beta Was this translation helpful? Give feedback.
and others
-
The base model in Frigate+ has now been updated to 2024.2. All model requests going forward will use this version. In order to get a new model, you will need to submit a model request.
The 2024.2 base model update was focused on reducing known false positives by selecting the highest scoring false positives that were reproducible with the 2024.1 base model. To date, nearly 500,000 false positive examples have been submitted by users. The previous base model reproduces 14% of these false positives and the new base model reduces that to 7%. One of the most common false positives was a person being identified as a dog while bent over, so hopefully that should be less common going forward.
I want to reiterate that you should not report false positives for objects that are in locations you want to ignore. You aren't likely to be able to convince your model that a person outside of your zones is not a person at all. This also makes it difficult for me to find good false positive examples to improve the base model.
I wanted to expand the label set with this release, and I have spent a significant amount of time adding labels to the existing base model training set already. However, I underestimated how significant the impact of adding new labels would be on the entire workflow. I am now working on the following updates which I believe are necessary for the an expanded label set:
Once these are done, adding additional labels will be much more straightforward.
In addition, I have almost finished refactoring the annotator in Frigate+ into a separate open source component library so it can be integrated directly into Frigate.
Beta Was this translation helpful? Give feedback.
All reactions