0.17.0 Release Candidate 3 #22074
Replies: 15 comments 42 replies
-
|
Does anyone use renovate for frigate updates and have a working config that includes beta and rc releases? I've tried so many things but can't seem to dial it in. |
Beta Was this translation helpful? Give feedback.
-
|
I have to say I am impressed with the improvements in RC3. After I was kicked off here I've forked Frigate to update all supporting structures so it will work with my new Blackwell GPU (SM 121, RTX 50xx), that is:
Each compile takes ~2.5 hours, so it is a long time between iterations, and I need 200GB of disk. (It's a VM) No love yet. I think I'll call it Cormorant. |
Beta Was this translation helpful? Give feedback.
-
|
Attached is my ORT (ONNX Runtime) wheel factory build with the latest supporting everything, for my AMD system with nVidia RTX 5050 Blackwell GPU, Debian 13.3, and Podman (not Docker, so no root nothin', even inside the containers). Still fussing with the cormorant (nee frigate) build, little TypeScript incompatibilities. ✔ Proper caching ✔ Podman-safe ✔ Minimal runtime Blackwell GPUs (SM 120, RTX 50xx)
Note I've set --parallel 5. When set to the 10 cores in the VM, the build crashes OOM. (16GB RAM -- I'm not giving it more) I've split ORT out as it is a long build, so I can get faster build iterations for cormorant. If no valued containers exist (start fresh): {file removed} |
Beta Was this translation helpful? Give feedback.
-
|
Huh. "Pull request creation failed. Validation failed: must be a collaborator" // Old // New Repeat for button and a elements. This makes the code fully TypeScript 5 / React 18 compatible. |
Beta Was this translation helpful? Give feedback.
-
|
TS5 hurts in 0.17 RC3
If going to TS5 these should be corrected. This is not a criticism; TS5 is demanding. If my suggestions aren't interesting or are ignored I can just patch Frigate only in my own build as I come across them. I'll understand. |
Beta Was this translation helpful? Give feedback.
-
|
Well, no clicks, no responses, it looks like I am just talking to myself. |
Beta Was this translation helpful? Give feedback.
-
|
SHE'S SINGIN'!! On all 16 cylinders! Now I have to actually learn Frigate. |
Beta Was this translation helpful? Give feedback.
-
|
I would like to suggest a modification to the documentation regarding hardware acceleration. I use a mini PC with an i3 1220p. A 12th generation processor and GPU. The documentation indicates a preset selection decision based on the CPU generation, however, in some cases, like mine, which uses XE architecture, the correct choice is QSY. According to the documentation, I was using VAAPI and was getting many ffmpeg errors and searching for answers in the wrong causes. The suggestion is to guide users to research the GPU architecture and not just look at the version. I believe the biggest problem lies in this intersection between 12th and 13th generation versions. |
Beta Was this translation helpful? Give feedback.
-
|
Anything after Beta 2, I get these errors on startup and facial recognition doesn't work. |
Beta Was this translation helpful? Give feedback.
-
|
understood. is there any way to identify the corrupt image? |
Beta Was this translation helpful? Give feedback.
-
|
Try disabling detect in the config. For some reason right at startup an object is being marked as stopped detected. Can you share your full config? |
Beta Was this translation helpful? Give feedback.
-
|
so, I went back to using |
Beta Was this translation helpful? Give feedback.
-
|
Not really, they logs have nothing to do with face recognition, it's an issue with the embeddings model. It is odd that seems to make a difference but really just makes for more questions. |
Beta Was this translation helpful? Give feedback.
-
|
so, I am back with everything working. I think it totally had to do with sitting in front of one of my cameras while stopping/starting frigate. Once I went into another room, everything is starting and working as expected including no more ffmpeg errors. Crazy that likely I was sitting in the same chair every time I updated? Yet rolling back, likely in the same spot, didn't cause issues? Anyway, I am in |
Beta Was this translation helpful? Give feedback.
-
|
Hey all, quick question concerning the Classification Models, when i create a New Classification for:
When I then try select the car it won't show in the list of potential pictures. Is there a way to "Upload" a picture so the system knows what I want. Or is that possible to select from the records which picture or in this case car I need. Maybe it would be possible to select it from "Tracked Object Details" Or did I miss something ? |
Beta Was this translation helpful? Give feedback.
-
|
From the docs:
|
Beta Was this translation helpful? Give feedback.
-
|
Thanks ! As always RTM = => Read The Manuel aka Doc. I did miss that part. |
Beta Was this translation helpful? Give feedback.
-
|
Hello, Issues :
Suggestions :
|
Beta Was this translation helpful? Give feedback.
-
|
Thanks for the feedback. The new custom roles feature has been thoroughly tested by a number of users during the beta period, and we never had any reports of this. I also cannot reproduce the issue of needing to refresh the page to see updates. I created a new role, and the roles page updates immediately on save with the new role I created. I can also edit cameras in that role, save, and I immediately see the updates. Can you give exact steps to reproduce your issue? |
Beta Was this translation helpful? Give feedback.
-
|
Thanks for the quick reply ! However, the rest is still valid : speaker button not on UI + “Audio is not available for this stream” + 403 errors in network tab for calls to https://192.168.x.x:8971/api/go2rtc/streams/cam_XX_main with response : {"detail":"Access denied to camera 'cam_XX_main'. Allowed: ['cam_XX']"} I just tried something in the config :
I hope those details help. |
Beta Was this translation helpful? Give feedback.
-
|
Thanks, your first issue was likely a cache issue. The second issue is something that we saw early on in development and was fixed, but maybe a later change caused a regression. We'll see if we can reproduce. |
Beta Was this translation helpful? Give feedback.
-
|
I was able to reproduce your issue, which was slightly different than the one that I had fixed a number of months ago. The changes are in #22097 |
Beta Was this translation helpful? Give feedback.
-
|
Lovely RC. Some feedback; I wish there was an option to multi-select snapshots for training. Currently you have to click every individual image and select the corresponding result for training. Selecting multiple and then assigning the correct object once would be a lot quicker. Also, when editing a classifiction state, you can't add another camera and/or change the area for monitoring (in case a camera FOV gets changed slightly because of cleaning it or something. Thank you! |
Beta Was this translation helpful? Give feedback.
-
|
See this tip on why multi-select/bulk selection is not implemented: #21374 |
Beta Was this translation helpful? Give feedback.
-
|
Thank for the read! I understand now. :) |
Beta Was this translation helpful? Give feedback.
-
any reason ??
|
Beta Was this translation helpful? Give feedback.
-
|
2026-02-23 19:53:57.924596357 [2026-02-23 19:53:57] ffmpeg.entrence.detect ERROR : [aost#0:1/aac @ 0x55d78d19d4c0] Non-monotonic DTS; previous: 22740094, current: 129016; changing to 22740095. This may result in incorrect timestamps in the output file. |
Beta Was this translation helpful? Give feedback.
-
Beta Was this translation helpful? Give feedback.
-
|
It looks like your cameras are sending corrupted data to Frigate. Are these wifi cameras? |
Beta Was this translation helpful? Give feedback.
-
|
yes it WIFI camera and yes my bad its from encoding issue due to low VRAM spike.. |
Beta Was this translation helpful? Give feedback.
-
|
I am not sure if it is just coincidence or not, but after updating from RC1 to RC3, my person detection isn't really happening at all. It is missing heaps of people for my street cameras. I am currently using GTX 1070 after generating the yolov9 model in the instructions. This is how it is in the config I have object filters set to this on two cameras. Camera 1 Camera 2 I guess a good question would also be, is this the best model to use for that card? |
Beta Was this translation helpful? Give feedback.
-
|
Is there a chance that the changes to the ONNX Runtime have also stopped support for the 1070 to correctly be used the detector? My system isn't detecting correctly at all |
Beta Was this translation helpful? Give feedback.
-
|
If it did it would error out when loading the model, not silently miss objects The changes also came well before RC |
Beta Was this translation helpful? Give feedback.
-
|
Thanks @NickM-27 , I was thinking I would see errors as well. It has just gone almost silent. Very few detections are showing up, but most aren't. Which I know must be wrong, because the same people walk past at the same time, the detections just stopped... I might try my Coral and see what happens |
Beta Was this translation helpful? Give feedback.
-
|
Coral is detecting again. I am using the new yolo v9 coral model yolov9-s-relu6-best_320_int8_edgetpu.tflite I might try the 1070 when I have more energy or just leave this... If anyone else starts to have the same issue, let me know and I will try sooner |
Beta Was this translation helpful? Give feedback.
-
|
Desktop 1) laptop 2) mobile ui 3) ALERTS sometime not showing. This is happening since long time not only in 0.17 where to see what is the issue does it come in log?? |
Beta Was this translation helpful? Give feedback.
-
|
Check the network tab in browser tools |
Beta Was this translation helpful? Give feedback.
-
|
is it here to check ? |
Beta Was this translation helpful? Give feedback.
-
|
yes, you'll want to check when some are missing though |
Beta Was this translation helpful? Give feedback.
-
|
How far are we from the 0.17 release on :latest? I am itching to upgrade to 0.17, but not sure if I should jump into 0.17 RC3. |
Beta Was this translation helpful? Give feedback.
-
|
Like every RC we're just waiting to see if new issue come up. |
Beta Was this translation helpful? Give feedback.
-
|
Gave it a try, runs fantastic. |
Beta Was this translation helpful? Give feedback.
-
|
I pulled the trigger and upgraded. Running without issues. Thanks for the great work! |
Beta Was this translation helpful? Give feedback.
-
|
I'm currently encountering an error when disabling the LPR feature in the UI: 
Below is my configuration file: |
Beta Was this translation helpful? Give feedback.
-
|
I finally found out it was because I had set lpr.enabled: true on one of the individual cameras. This isn’t a serious bug, and it will probably be fixed when the settings page is refactored in a future release. |
Beta Was this translation helpful? Give feedback.
-
|
I want to configure a state classification for a garage door. It is a flip open garage door. The states are open and closed. How do I handle the inbetween states when it is the process of opening or closing. Should I train it on images where the garage door is partially open? If I define a cutover point in the opening or closing process, the appearance of the door will be very similar at that cutover point for both states. |
Beta Was this translation helpful? Give feedback.
-
|
I think you can kind of choose how you want it to work based on your needs. For me personally on my gate I made any image where the gate is not closed (ie there is a gap between the gate) marked as open. |
Beta Was this translation helpful? Give feedback.
-
|
Is the images for state training fixed to square ratio? I have a side angle view of the garage door, but to include the full door height I have to include a lot of unrelevant space. I was thinking of using only a small square section at the bottom edge of the one side of the garage? |
Beta Was this translation helpful? Give feedback.
-
|
yes it must be a square, the irrelevant pixels shouldn't matter as it generally will learn what is changing with the states |
Beta Was this translation helpful? Give feedback.
-
|
Hey all, Small question in the "explore" part. When you review a clip there used to be a button to switch to Full screen. This button is no longer present. In order to see it in full screen, the ... dots needs to be clicked and then select "View in History". Would it be an idea to bring the full screen button back in 0.18 as for 0.17 I understood no new function will be implemented. |
Beta Was this translation helpful? Give feedback.
-
|
Is it still possible to submit a detected face for classification from the explore screen? I seem to recall there being a button for this previously but there no longer seems to be |
Beta Was this translation helpful? Give feedback.
-
|
No it's not possible, we got a lot of feedback that users found this confusing, so the face training is kept in the face library UI |
Beta Was this translation helpful? Give feedback.
-
|
Ah, OK. Must admit I found it a pretty useful feature when trying to find training images for new faces where there weren't many examples in the face library, but I appreciate you probably get feedback both ways on any such features |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
Beta docs: https://9eaa7bfe.frigate.pages.dev/
Images
Changes since RC2
Full changelog
Major Changes for 0.17.0
Breaking Changes
There are several breaking changes in this release, Frigate will attempt to update the configuration automatically. In some cases manual changes may be required. It is always recommended to back up your current config and database before upgrading:
frigate.dbfilegenaiconfig now only configures the provider. Other fields have moved underobjects -> genai. See the new GenAI documentation.record -> continuousandrecord -> motionare separate config fields. See the examples in the documentation.smallmodel, which performs well on both CPU and GPU. Thelargemodel is the same as 0.16's and is not as accurate as the upgradedsmallmodel in 0.17. Uselargeonly if you live in a region with multi-line plates and you are having issues detecting text on them with thesmallmodel.detectresolutionwidthandheightfor cameras in your config if Frigate hangs on startup.exec,expr, andechosources for go2rtc are now removed by default to reduce the security risk if an attacker has access to the configuration. This can be disabled using an environment variableGO2RTC_ALLOW_ARBITRARY_EXECA separate configuration for this for HA addon users will come in a later beta. See the documentation.New Features
Frigate 0.17 introduces several major new features.
Classification Model Training
Frigate 0.17 supports classification models in two separate types: state classification and object classification. These models are trained locally on your machine using
ImageNetviaMobileNetV2.State Classification
State classification allows you to choose a certain region of camera(s) with multiple states, and train on images showing these states. For example, you could create a state classification model to determine if a gate is currently open or closed.
See the documentation.
Object Classification
Object classification allows you to choose an object type, like
dog, and classify specific dogs. For example, you can train the model to classify your dogFidoand add a sub label, while not labeling unknown dogs. Another example would be classifying if a person in a construction site is wearing a helmet or not.See the documentation.
Custom Viewer Roles
Frigate 0.17 now has the ability to create additional viewer user roles to limit access to specific cameras. Users with the
adminrole can create a uniquely named role from the UI (orauth --> rolesin the config) and assign at least one camera to it. Users assigned to the new role will have:viewerrole has access to (Live, Review/History, Explore, Exports), but only to the assigned camerasSee the documentation.
Review Item Summary with GenAI
Frigate 0.17 supports using GenAI to summarize review items. Unlike object descriptions which add a searchable description, review summaries have a structured output that instruct the AI provider to generate a title, description, and classify the activity as dangerous, suspicious, or normal.
This information is displayed in the UI automatically making it easier to see when activity requires further review and easier to understand what is happening during a particular video segment.
See the documentation.
Semantic Search Triggers
Triggers utilize Semantic Search to automate actions when a tracked object matches a specified image or description. Triggers can be configured so that Frigate executes a specific actions when a tracked object's image or description matches a predefined image or text, based on a similarity threshold. Triggers are managed per camera and can be configured via the Frigate UI in the Settings page under the Triggers tab.
See the documentation.
Object Detector Improvements
Frigate 0.17 brings performance increases for many detectors as well as support for new object detection hardware.
Nvidia GPU Performance
Support for Nvidia GPUs has been enhanced by implementing CUDA Graphs. CUDA Graphs work to reduce the involvement of the CPU for each inference, leading to faster inference times and lower CPU usage. CUDA graphs do have some limitations based on the complexity of the model, which means that YOLO-NAS, Semantic Search, and LPR models are not accelerated with CUDA Graphs. They will still continue to run on GPU as they did before.
Intel OpenVINO
Frigate 0.17 supports running models on Intel NPUs, for many models performance on NPU is similar to GPU but more efficient, leaving room to run more enrichment features on the GPU.
OpenVINO has also had many optimizations put in place to reduce memory and CPU utilization for object detection.
RKNN
Frigate 0.17 brings several improvements to RKNN platform including:
largemodel sizes.Apple Silicon
Frigate 0.17 supports running object detection on Apple Silicon NPU. This is provided through the Apple Silicon Detector which runs on the host and connects via IPC proxy to Frigate, providing fast and efficient inferences when run within the same Apple device.
See the documentation.
YOLOv9 on Google Coral
Frigate 0.17 supports running a quantized version of
YOLOv9on Coral devices, bringing improved accuracy over the defaultmobiledetmodel. Note that due to hardware limitations, only a subset of the objects on the standard COCO labelmap is included. YOLOv9 Frigate+ models are not supported on Coral at this time.See the documentation.
New Community Supported Detectors
Frigate 0.17 has community support for several new object detectors:
Frontend Improvements
In addition to supporting the new features, the frontend has many improvements.
Detail Stream
History view in 0.17 supports an additional view mode, Detail. This mode shows a card for each review item, and expanding a card reveals all tracked objects and their lifecycle events. Selecting any lifecycle event seeks the video to that exact timestamp. You can also overlay a tracked object's path on the video to help with debugging.
Redesigned Tracked Object Details pane
The Tracked Object Details pane in Explore has been redesigned to streamline the layout and consolidate related information. The Object Lifecycle tab is now the Tracking Details tab, which displays video overlays of the tracked object instead of static images, giving a clearer and more intuitive view of its activity.
Revamped Settings
Frigate 0.17 has a revamped Settings menu with a sidebar that categorizes the available options. This brings more scalability which will make it easier to support full UI configuration in a future version.
NOTE: The Debug view has been moved to the single camera Live view instead of Settings. Access the Debug view by enabling the switch under the Live view settings (cog icon) menu.
Add Camera Wizard
Frigate 0.17 supports adding camera via the UI without manually modifying your configuration file. When installing and starting Frigate for the first time, the main dashboard will include a button to start adding cameras via the Wizard.
Access the Wizard from the
Cameras --> Managementpage in Settings.Update Without Restarting
Frigate 0.17 supports saving many more features dynamically. Cameras, zones, and masks will not require a restart to take effect when saved through the UI. More will come in future versions.
Configuration Safe Mode
If an invalid configuration is detected, Frigate will enter safe mode and highlight the location of the issue. While in safe mode, the frontend is limited to the configuration editor, making it easy to correct the problem directly in the UI without needing an external file editor.
Other Notable Frontend Improvements
Other Backend Features and Improvements
Audio Transcription and Analysis
Frigate 0.17 supports fully local audio transcription using either
sherpa-onnxorfaster-whisper. The single camera Live view in the Frigate UI supports live transcription of audio for streams defined with the audio role, and anyspeechevents in Explore can be transcribed and/or translated through the Transcribe button in the Tracked Object Details pane.See the documentation.
Process and Efficiency Improvements
Frigate 0.17 uses the forkserver spawn method, this allows for better segmented memory control and better process management. Some processes are also started with lower priority, allowing the most important processes to have more CPU time when it is required.
Review Item Improvements
Review items have been refined to behave more intuitively:
Revamped stationary object tracking. Stationary object tracking has been enhanced to use new features to reduce incorrectly marking objects as active:
Smarter handling of loitering objects. Stationary behavior is now dynamic based on object type. Objects that are normally stationary for long periods (e.g., cars) will no longer keep a review item active indefinitely when stopped inside a loitering zone. Objects that are not expected to remain still (e.g., people) will continue the review item as long as they stay within the zone.
Severity-based review item cutoff. Review items now end when a higher-severity event (such as an
alertfor arriving home) finishes. Ongoing lower-severity motion (e.g., passing cars) will no longer keep the higher-severity review item alive. In these cases, thealertends and a newdetectionreview item begins immediately.Enrichment Improvements
-,deviceconfig option. This is useful in cases when multiple GPUs are available. DocumentationOther Improvements
networking -> ipv6 -> enabled. DocumentationThis discussion was created from the release 0.17.0 Release Candidate 3.
Beta Was this translation helpful? Give feedback.
All reactions