0.16.0 Beta 3 #18501
Replies: 33 comments 209 replies
-
|
|
Beta Was this translation helpful? Give feedback.
-
|
You should clear your browser cache. |
Beta Was this translation helpful? Give feedback.
-
|
The documentation gives a good general explanation of scoring for a set of faces but maybe some clarification on the formula with some examples? I also note that when the GROUP gets an Unknown marking even when there where matches it doesn't have a score.. It would be potentially useful to know how close to the current threshold the group was. Reprocessing individual faces also does not appear to refresh the group score. Overall this is really starting to get dialed in, false positives seem to be down, however I ask about the scoring as I have one set that had a few strong matches but failed to "MATCH" as a group so I was just trying to resolve how the formula worked without going into the code to figure it out. |
Beta Was this translation helpful? Give feedback.
-
|
Anyone else noticing this? What are your min scores set to? |
Beta Was this translation helpful? Give feedback.
-
|
I've only seen it when a person is wearing a hat and it gets the ear included, I don't see back of the head faces |
Beta Was this translation helpful? Give feedback.
-
|
Beta Was this translation helpful? Give feedback.
-
|
In your first example there is enough face there that I have gotten a match also your score is 77%, mine is 74%.. What do you have your min_score set to for face? I have been going up and down between 72%-81%, however at 75% and higher I get very few BAD faces but I also have far fewer face matches in general or only samples or 1-2 faces in some events. |
Beta Was this translation helpful? Give feedback.
-
|
I just use the default 0.7 |
Beta Was this translation helpful? Give feedback.
-
|
On my side disabling the user setting "Automatic Live View" doesn't seem to show any effect. |
Beta Was this translation helpful? Give feedback.
-
|
Are you using a camera group? Have you overridden the settings there? https://deploy-preview-16390--frigate-docs.netlify.app/configuration/live/#streaming-options-on-camera-group-dashboards |
Beta Was this translation helpful? Give feedback.
-
|
I've fixed a bug that prevented the switch in Settings from working correctly in #18500. It will be available in the next beta. |
Beta Was this translation helpful? Give feedback.
-
|
Thanks, MQTT snapshots appear to be working again! |
Beta Was this translation helpful? Give feedback.
-
|
I actually just raised an issue with beta3 that my snapshots are not working! I think they were working with 0.15, but not anymore with 0.16-beta3. Edit: Yeah, when updating to beta THREE it works.... |
Beta Was this translation helpful? Give feedback.
-
|
does it support the new vision model from Qwen? https://ollama.com/library/qwen2.5vl |
Beta Was this translation helpful? Give feedback.
-
|
Frigate just hands off requests to the Ollama API. If you've configured GenAI correctly, there's no technical reason why any vision-enabled model with Ollama wouldn't work. You should watch the Ollama debug logs and/or enable embedding maintainer debug logs with Frigate: |
Beta Was this translation helpful? Give feedback.
-
|
Can confirm that qwen2.5vl works well for me |
Beta Was this translation helpful? Give feedback.
-
|
Beta 3 fails to start with ONNX/ROCM |
Beta Was this translation helpful? Give feedback.
-
|
I can see this file is correctly included. Not sure why this error would happen for some but not others. |
Beta Was this translation helpful? Give feedback.
-
|
Repo: ghcr.io/blakeblackshear/frigate:0.16.0-beta3-rocm The ONNX runtime does not appear to have the MIGraphX provider ( Failed to load library libonnxruntime_providers_migraphx.so) |
Beta Was this translation helpful? Give feedback.
-
|
Same here. Fresh install. |
Beta Was this translation helpful? Give feedback.
-
|
Yeah, looks like the 6.3.3 wheel specifically is missing it. In testing I have found that rocm performs better anyway, so it will be preferred in the next beta |
Beta Was this translation helpful? Give feedback.
-
|
Another observation: |
Beta Was this translation helpful? Give feedback.
-
|
This is a bug with Intel GPU tools. A reboot would likely fix it again for a short period. |
Beta Was this translation helpful? Give feedback.
-
Beta Was this translation helpful? Give feedback.
-
|
Thanks. I still don't see any issues of concern. Earlier, you said:
This seems quite odd and might indicate missing key frames. You could try inspecting the key frames through: Since we cannot reproduce this and other users are also saying beta 3 is working well, I'm not sure what can be done at this point other than to see if other users report a similar issue and then see what the commonalities are. |
Beta Was this translation helpful? Give feedback.
-
|
I appreciate the time spent looking at this. As for the odd times shown in the video player, I mentioned that because it did seem odd. It happens with clips downloaded from both Beta 2 and Beta 3. I don't know if it's relevant, but I just tried using the Export feature to get the same time period (with times from the Lifecycle screen), and that resulted in a video that is about 1.7MB larger, includes a bit more time at the beginning (6 additional seconds for 36 second total video even though the times selected are only 25 seconds apart), and no oddities with the video player timeline (Totem on Ubuntu). Downloading the clip from event again (in case something glitched in the download), Totem still shows a 20 second countdown but is able to play the full 30 seconds. Totem consistently shows a playtime that is an exactly multiple of 10 and shorter than the downloaded clip even though it plays the full clip (another I just compared is actually 17 seconds of video but Totem shows it as 10 seconds even though it can be coerced to play the additional 7 seconds after its thumb reaches its artificial end). In each case, I'll try again when the next beta release comes out. My five cameras, BTW, all show this behavior except the one that uses H.264 (because it doesn't support H.265). That camera shows clips in the web UI that start almost precisely matching the detection event from the Lifecycle (within a frame or two) So I suspect that's the underlying issue. edit Just saw the thing about running |
Beta Was this translation helpful? Give feedback.
-
That's an important and interesting detail. Earlier you noted:
Does the behavior improve for new recordings if you set a keyframe interval for the record stream? |
Beta Was this translation helpful? Give feedback.
-
|
So to set a keyframe interval, I had to switch off the “Smart Codec” feature. That does seem to be the culprit, though. The web UI clips now start at a reasonable place. The recordings are 6x the file size they used to be, though, which makes the web UI stutter at segment boundaries. So I'll have to decide between reducing the quality or dealing with unusable web UI clips. Thanks for helping me track that down. |
Beta Was this translation helpful? Give feedback.
-
|
Great, glad we've found the issue. "Smart" codecs remove keyframes, which are extremely problematic for live streaming. We've documented this and provided some additional data here: https://deploy-preview-16390--frigate-docs.netlify.app/configuration/live/#camera-settings-recommendations |
Beta Was this translation helpful? Give feedback.
-
|
I'm attempting to use an AMD GPU to perform embeddings, using the rocm image. I now see this error. |
Beta Was this translation helpful? Give feedback.
-
|
For anyone else with this issue, I had to set |
Beta Was this translation helpful? Give feedback.
-
|
Thanks for updating with your resolution, that is interesting. I will investigate to see if we can include this by default |
Beta Was this translation helpful? Give feedback.
-
|
I have no idea what it does, but I found several GH discussions about setting it to resolve the error above. As in: jax-ml/jax#14582 |
Beta Was this translation helpful? Give feedback.
-
|
it just tells the model runner to use the default convolution method instead of trying to calculate the best one, basically saves on CPU usage and I guess prevents it from running into errors on certain models. The downsize is potentially slower inference but in my testing it did not really change anything for object detection models so I went ahead and added it as a default env var for the next release. |
Beta Was this translation helpful? Give feedback.
-
|
I have moved all of my cameras to full / primary stream resolution now for detect. Most of my face and plate detection issued are gone now. It can clearly read plates way more often and it can differentiate family members MUCH more reliably. Good clear sample faces for training are also more frequent. Probably going to move my frigate box to more powerful hardware in the future.. The 20-30% CPU camera warnings are little annoying and if ALL cameras have too much activity it does drive the CPU use up a lot, however with 4 cores and 8 cameras that isn't too much of a problem. |
Beta Was this translation helpful? Give feedback.
-
|
In general I don't see anything remotely comparable to your example until 1000px processing / compression etc are always factors. mostly noting that face detection "DID NOT WORK" on ANY of my cameras at 500px. |
Beta Was this translation helpful? Give feedback.
-
While a higher pix default would mean closer detections it would still WORK on higher end cameras. The default should work out of the box for "MOST" configurations and the user and adjust UP or DOWN but it should "work" out of the box for most configurations both high and low end. The Reolink Doorbell and a lot of the Tapo cameras seem to be fairly commonly used, I don't think the feature will work AT ALL with the current default. The Reolink doorbell and most doorbells have horrible compression and people generally walk toward the camera from a distance, would be a shame for the samples to fill up before the person even is clearly identifiable. |
Beta Was this translation helpful? Give feedback.
-
|
I think it would be more problematic to have such a high min area that would potentially lead to very few detections, especially when running the separate face detection. If we get feedback that faces are getting stuck with bad examples and no examples of a person being closer are included then we can look at changing the defaults but so far we haven't seen that feedback much |
Beta Was this translation helpful? Give feedback.
-
|
People also like to turn around in my drive way. I will certainly agree that my experience is bias. Would be much better if we got samples from several users. However looks at the few reddit and "its not working" threads so far lower quality cameras seem to be common. Also my two suggestions are actually LOWER than my actual settings. of 2300 for face and 3000 for plate currently.
This one was at a lower detect resolution |
Beta Was this translation helpful? Give feedback.
-
Absolutely, again just sharing my experience and I only have your camera examples to go off. I do have 3 different brands and various "qualities" of cameras however so I was curious how the default could be so low. My Dahua is only 1080p gets less blurry results but is very grainy, still can't really use faces or plates it gets at the defaults and it is in my garage facing the garage door when the car pulls in. |
Beta Was this translation helpful? Give feedback.
-
|
Is there any indication when 0.16 might get elevated away from Beta to production? Also keen to explore LPR and facial recog of course! |
Beta Was this translation helpful? Give feedback.
-
|
No, we'll have at least one more beta and then a release candidate as well |
Beta Was this translation helpful? Give feedback.
-
Thx, I'm off to see how to switch to Beta then 😄 |
Beta Was this translation helpful? Give feedback.
-
|
Still getting plenty of these unable to create snapshot warnings in my log - sometimes once a minute. Have tried numerous things, including increasing the /tmp/cache size, disabling snapshots on two of my cameras that have produced issues in the past (which was fixed by adding the #media=video in the go2rtc config), and upgrading my hardware from an intel 4770 cpu machine to a 9700 cpu machine (moving from i965 to iHD vaapi drivers). Have also tried using openvino as a detector instead of a coral just as a random experiment. Never saw these errors in 0.15 but have seen them in all three 0.16 betas. My config is: Docker config is (including the enlarged tmpfs): |
Beta Was this translation helpful? Give feedback.
-
|
Go ahead and open your own support discussion with screenshots of your system metrics and camera metrics pages. |
Beta Was this translation helpful? Give feedback.
-
|
since beta2 it is not stable - always stop working siddenly =( beta1 works fine exept PTZ controls - they dont work at all =( |
Beta Was this translation helpful? Give feedback.
-
|
The PTZ changes from beta 1 to beta 2 were not trivial. It would be helpful to see a full copy of the logs from Frigate startup to the time it becomes unresponsive, not just a snippet of nginx requests. |
Beta Was this translation helpful? Give feedback.
-
Beta Was this translation helpful? Give feedback.
-
|
Can you confirm your issue disappears if you revert to beta 1? You will need to restore from your backup. |
Beta Was this translation helpful? Give feedback.
-
|
yes, the issue doesnt exist in beta1. |
Beta Was this translation helpful? Give feedback.
-
|
We'll need more information to be able to help further. Go ahead and open up your own support discussion with your full config and logs and describe exactly what you're experiencing and how you reproduce it. |
Beta Was this translation helpful? Give feedback.
-
|
Two interesting bugs.
|
Beta Was this translation helpful? Give feedback.
-
not sure what you mean by "taking blurry snapshots", it just looks for faces in the people that are detected. If it is blurry that is a characteristic of the stream quality, and not something Frigate can work around. |
Beta Was this translation helpful? Give feedback.
-
The initial/default min_area setting is 500 try increasing it to 1000 or 2000 and see if the images improve, if they do adjust up or down as needed. 500 was completely unusable on my cameras. |
Beta Was this translation helpful? Give feedback.
-
|
I think part of my confusion here is that the blowed up face photo is different than the snapshot that's shown when you click on the review item. I've seen 6000px face items that are showing 0% face score. |
Beta Was this translation helpful? Give feedback.
-
The snapshot is just ONE instance during object tracking multiple face samples can be taken and if there is one it isn't always the same moment as the event snapshot. If there are multiple people in a room, close to each other / overlapping from the cameras view, sometimes the wrong face gets sampled. |
Beta Was this translation helpful? Give feedback.
-
|
Would you all be open to a config variable to disable CPU warnings? Any user running docker in a VM thru a hypervisor like Proxmox sees them constantly and it's not an accurate portrayal of the real usage since hypervisors usually dynamically allocate resources on demand between all VMs and the host. |
Beta Was this translation helpful? Give feedback.
-
|
This was discussed in the previous beta post, we don't want to make the levels configurable as that restricts and adds lots of complexity to it. I'm also hesitant to add options to disable as we either need to have many options for each feature / metric or risk a user disabling everything and missing out on important information. |
Beta Was this translation helpful? Give feedback.
-
|
No, working around hypervisor issues isn't something we want to implement generally, and we've tried to avoid putting UI-related options in the config. Additionally, there are likely times when a warning message would actually be important to see. We have plans to make improvements in this area. A future version of Frigate will include a "potential problems" pane (think Home Assistant's "Repairs" pane), which could include the ability to dismiss the messages. For now, you could use a browser extension to hide the div with the CPU warnings, if it's really that much of an annoyance. |
Beta Was this translation helpful? Give feedback.
-
High CPU usage reported inside VM while PVE shows low CPU usage Check that your guest VM CPU type is set to host to ensure all host features are being passed through to the guest. I don't actually see a substantial "per core" difference in the CPU use reported. I do get the CPU warnings now but that is because I started running Full resolution for all of my detect feeds. Before that I had no CPU warnings even on an old 7th gen CPU. Edit: to further clarify host is the best setting "UNLESS" you have a cluster and need VMs to live migrate, however that is unlikely as you probably are doing GPU passthrough which can't live migrate anyway. |
Beta Was this translation helpful? Give feedback.
-
|
Amazing release so far, it is awesome that all of this works locally. I am wondering if there are any recommendations for good cameras that could be used for dedicated LPR, I know I can look elsewhere but wasn't sure if there was a Frigate recommendation |
Beta Was this translation helpful? Give feedback.
-
|
No particular recommendations from us devs, but the users at the ipcamtalk forums usually are really helpful. |
Beta Was this translation helpful? Give feedback.
-
|
When adding or removing pictures from training there seems to be a background process that runs before the model updates and clicking on reprocessing of photos will ACTAULLY use the updated info. Would it be possible to get a popup or something that said something like "Face model updated for [Name]" |
Beta Was this translation helpful? Give feedback.
-
|
Is there some python I could run against an image to TEST the OCR results with different enhancement levels or different sample images? I am getting infrequent matches but I can clearly read the plates or partial plates easily from the plate DEBUG so I am not sure what the is better or worse from the OCR software's point of view. Having a DEBUG screen where you could play with the settings and immediately see the results to tune for each camera would be awesome. I don't think OCR cares for color much and needs ULTRA high contrast I have used whiteboard capture tools at work that do a great job of flattening "photos" into two tone high contrast . Looks like this is possible with openCV, so I am curious if I could pre-process some images then TEST it with some CLI using frigates setting to see if it helps. https://santhalakshminarayana.github.io/blog/whiteboard-image-enhancement-opencv-python Same with de-skew. some of the plates look a bit off angle due to my camera angle AND the approach of the car https://stackoverflow.com/questions/78337034/deskew-image-using-opencv-in-python |
Beta Was this translation helpful? Give feedback.
-
|
Thanks for the feedback. In order for the cameras to do its primary job of capturing the full front yard I can probably only lower it a bit it is almost perfectly centered on the property. I may produce some samples from my garage cam that is more dead on to the car entering the garage.. it works maybe 3/5 times of the car entering and leaving the garage. |
Beta Was this translation helpful? Give feedback.
-
|
Lowering it would likely help. I know that @NickM-27 also has a similar wide angle camera, but it's positioned much lower than yours and picks up plates in his driveway very well. |
Beta Was this translation helpful? Give feedback.
-
|
Lowered the camera about 1.5 feet. Had to lower recognition_threshold: a little and enhance a little but it seems to have worked for the last few car passes. Since there is no UI and the DEBUG level is very noisy any chance the "Detected text" and "Plate comparison" verdicts could be moved up to INFO level logging? This seems like a resonable level of detail to always log and would allow for more troubleshooting. |
Beta Was this translation helpful? Give feedback.
-
|
Frigate doesn't log anything else related to object detection at the INFO level, so this would be a bit odd to do. I'd suggest just using grep with the log, something like |
Beta Was this translation helpful? Give feedback.
-
|
@hawkeye217 |
Beta Was this translation helpful? Give feedback.
-
|
Regarding rocm and onnx, will onnx be on par with rocm performance wise ? Will you release YOLOv9 model in Frigate+ ? In any case, another great release coming, thank you devs ! |
Beta Was this translation helpful? Give feedback.
-
|
It is the same performance on my test device which is an integrated 780M GPU. In fact with beta3 the inference speed is actually better, 20ms with low GPU usage except under high activity. |
Beta Was this translation helpful? Give feedback.
-
|
One observation: Also when clicking on a face, I would expect to be bought to either the snapshot that face was taken from or the video exactly at the timecode that the face was taken from. That would allow me to quickly train difficult faces where you need context like clothes etc. Also a button "re-evaluate all" would be cool. |
Beta Was this translation helpful? Give feedback.
-
|
Ah, okay the delay explains it. Could we maybe get a litte green status update once retrained model is ready? |
Beta Was this translation helpful? Give feedback.
-
|
The model takes a few seconds to train based on the number of images (about 1 second per 100 images) so for the most part the delay should not matter. It is mostly just difficult because we don't want to re-train the model after every change is made since multiple changes may be made in a row, so really it is after the first reprocess that the model training happens. It definitely can be more clear but I think that is more of a rework of some of the UI vs a quick change. |
Beta Was this translation helpful? Give feedback.
-
That explains a lot. I just thought it took a long time and that I had tested too early. Turns out that was triggering it? Seems like a reprocess button on the training page and a warning that there are unprocessed changes would clear that up. |
Beta Was this translation helpful? Give feedback.
-
|
These parts of my comment may have gone unnoticed because they were added via edit:
Regarding the bulk reprocess faces part, Moreover I think the faces are limited to 100 and you said training 100 faces should be around 1 sec so inference should be even less than that. So it would really be negligible. One could even consider of reprocessing all faces automatically once an updated model becomes available. This would solve the whole issue of not knowing when to run training and knowing when the new model is available. I had a brief look at the code for the face recognition and have some improvement suggestions:
|
Beta Was this translation helpful? Give feedback.
-
If you can't recognize from the face who it is then it's probably not a good face to be using for training.
I think you misunderstand, bulk training is not not worth it. It is specifically discouraged and would lead to worse results. You should not be training that many images at once. For reference I have ~80 images for each person and I have stopped training entirely because so far face rec on my cameras is very accurate. The goal is not to pick a whole bunch of images and train all of them.
They are normalized |
Beta Was this translation helpful? Give feedback.
-
|
Having an issue: I have in config: Detection disable works fine by itself. When I also disable motion, one camera logs: The others, I simply get FFMPEG crashes unexpectedly. Turning back on motion detection, the errors go away. |
Beta Was this translation helpful? Give feedback.
-
|
We'll need to see a full copy of the logs. |
Beta Was this translation helpful? Give feedback.
-
|
I was working to remove all of the UI/GET log lines because some of it had sensitive data in it...then I decided to close the browser, restart the frigate container to capture logs without the UI/GET lines. After this, the error went away and I am not able to reproduce it again. Do you know if there is a way to suppress the HTTP logging? |
Beta Was this translation helpful? Give feedback.
-
|
you should just use the download log button in the UI, the logs are separated out between different components |
Beta Was this translation helpful? Give feedback.
-
|
Ah, that helps. Sometimes I "blind myself" to certain features. I'm glad I asked. This will make things easier than looking through the docker logs. |
Beta Was this translation helpful? Give feedback.
-
|
Where do I find the folder with the face training images? I though it was in Media, but it's not there. I'm running HAOS baremetal, with Frigate as an Add-on. |
Beta Was this translation helpful? Give feedback.
-
|
It is in the same media folder. You won't see it in the HA UI though because that comes from the integration |
Beta Was this translation helpful? Give feedback.
-
|
Ok, thanks. How do I delete faces I don't want? I can't see a way in the UI in the Face library tab, and looking at the Media folder, via Samba, as you say, it's not visible, although I could have sworn it was visible on Beta 1? |
Beta Was this translation helpful? Give feedback.
-
|
Right click the face |
Beta Was this translation helpful? Give feedback.
-
|
Right, got that bit now, thank you |
Beta Was this translation helpful? Give feedback.
-
|
I've just discovered, now that there is a right-click option, that you can also right-click on the "train" button and delete face labels, which is what I was after! |
Beta Was this translation helpful? Give feedback.
-
|
@hawkeye217 |
Beta Was this translation helpful? Give feedback.
-
|
Question from something odd I was seeing in debug view.. Should license_plate be persistently detected on a parked car? I was looking at debug and my parked car is correctly detected but I noticed that the license_plate would flash for a moment then disappear every few min/sec while looking at the parked car. |
Beta Was this translation helpful? Give feedback.
-
|
No, license plate is not tracked so it won't show consistently for a parked car. It only shows for one frame when the stationary check happens. |
Beta Was this translation helpful? Give feedback.
-
|
Sorry Team, It was not the LPR that was not working it face the face detection. Sorry about that. Here is the log |
Beta Was this translation helpful? Give feedback.
-
|
Thanks for confirming. This is an issue for builds beyond beta 3, and will be fixed in the next dev build. |
Beta Was this translation helpful? Give feedback.
-
|
Guess I can equate my result to small test sample size. Plate detection just sort of sometimes works. |
Beta Was this translation helpful? Give feedback.
-
|
I'm unable to get face recognition to work on my doorbell. It works fine on my other cameras. Would very much appreciate assistance to catch where I'm messing up on this camera. Previously with doubletake it worked pretty good with the person image being passed to doubletake for compreface review. I get zero face detections and zero to train on the doorbell camera. Originally, I tried to get it for frontyard, sidewalk and porch only (ignoring pavement) but at this point I don't care and would take anything. I may have really botched the config, I don't know. Any insight out there? Relevant pieces (I think) from my configuration file: A sample mqtt listener result: |
Beta Was this translation helpful? Give feedback.
-
|
I don't have any max_area definitions anywhere. I'm learning so much right now though as I didn't even think about debug and hadn't had a reason to mess with it so much previously. I didn't think at all about object masks being an issue because detections work. The object mask is roughly the same as the motion mask. Enabling "motion mask" blacks out more than half the screen. People are detected, but it's based upon the frigging legs to groin snapshot. I have a super large object mask because I'm on a cul-de-sac in a slightly elevated lot with a lot of visibility of the area and I get so much alerts from the sidewalk, cul-de-sac and neighbors across the street. So now I guess the question is how do I prevent detections/alerts in that area but still have the image available for face recognition. |
Beta Was this translation helpful? Give feedback.
-
|
I'm not talking about motion mask I am talking about an object mask, which I assume is the same thing you are referring to, this is almost certainly the source of your problem. You use required zones to define the areas that you care about. |
Beta Was this translation helpful? Give feedback.
-
|
Sure, I have zones setup and can get more in depth with them. Curiously, if an individuals lower half is in the "sidewalk/porch" zone and the head is in the "street" so to speak and not in the zone, will it still get a face read? I assume so, but before I do a big backspace and start over with it all wanted to ask. |
Beta Was this translation helpful? Give feedback.
-
|
you already have your zones set as required so you should be fine to just delete the huge masks, and face is not affected by zones in any way |
Beta Was this translation helpful? Give feedback.
-
|
Thanks man. I wouldn't have enough thought about this stuff without your help since all was working with doubletake via person mqtt. You are a hero. |
Beta Was this translation helpful? Give feedback.
-
|
I have been running Beta 1 and Beta 2 without any problems but with this Beta 3 I am having some trouble loading new recordings / events on one of my cameras. It'll never load these on any device - it'll just get stuck with the loading spinning icon. If you jump around on the timeline in the camera recording history this usually gets it to load ok 99% of the time. Existing recordings / events before the Beta 3 installation seem to mostly load fine, only a few out of many had an issue but I will monitor this. It seems to be just an issue with one camera but I am still testing. If it helps and/or related the camera with the issues is using Hikvision H.264+. The camera which is working fine is using Hikvision H.265+. They are quite different camera models apart from this difference. I will restart that camera now but I do not believe it to be an issue suddenly with this camera. I cannot see anything related in the logs, nor the Chrome developer console. Thank you. Edit: After turning off Hikvision H.264+ everything seems fine, but this did work fine before with all previous versions of Frigate so I thought I'd mention it. The camera restart alone didn't change anything and the camera was otherwise untouched. The camera running Hikvision H.265+ is still fine. Thanks |
Beta Was this translation helpful? Give feedback.
-
|
Keep in mind + variants of the encoding is not recommended as has been seen to cause issues |
Beta Was this translation helpful? Give feedback.
-
Ok I can try with it off but I just wanted to mention this as it's been smooth sailing before. The retention period it allows, at least to me without any visual quality trade-off is immense. Thank you |
Beta Was this translation helpful? Give feedback.
-
|
After turning off Hikvision H.264+ everything seems fine, but this did work fine before with all previous versions of Frigate so I thought I'd mention it. The camera restart alone didn't change anything and the camera was otherwise untouched. Thanks |
Beta Was this translation helpful? Give feedback.
-
|
Hello, I try this beta version an I have a big issue after enable LPR or Face Recognition. My HomeAssistant was stuck after play record or search events. All others containers works fined and I need to reboot HA container. I have encountered this problem 7 times in less than a week. I disabled LPR/FR and all works fine |
Beta Was this translation helpful? Give feedback.
-
|
What version of Home Assistant are you using? |
Beta Was this translation helpful? Give feedback.
-
|
The last stable version |
Beta Was this translation helpful? Give feedback.
-
|
There are known issues with HA 2025.5 and later with the Frigate addon. See home-assistant/core#145592 You may be running into this same issue. If so, users have reported that downgrading to 2025.4 fixes it. |
Beta Was this translation helpful? Give feedback.
-
|
It's strange, since I disabled LPR and facial recognition, I no longer have the problem. I didn't have the issue before the beta update either. I think there's a connection, because it's since this change that I've been seeing this behavior. We'll see if other people are affected. |
Beta Was this translation helpful? Give feedback.
-
|
Is there a way to exclude alerts with matching faces. For example, I don't want to be alerted to my family when whey are at a exterior door. I know you can use 3rd party tools to filter this out but can you do it natively in frigate, and if not will it be included in version 17? thanks |
Beta Was this translation helpful? Give feedback.
-
|
I don't think it really makes sense, downgrading a review item would add a lot of complication. And in general just because a known person or car is in the video doesn't mean there is not something to review. For example, you could walk inside and then someone could follow behind and open your car door. That also wouldn't stop the notifications coming when Frigate only detected person before the face is recognized, which could be confusing if you try to open the notification and it no longer exists. It's possible there could be an option to automatically mark it as reviewed. |
Beta Was this translation helpful? Give feedback.
-
|
Not sure if this is related to beta3 or if this is a general issue, but when I export clips, there are no thumbnails. Log shows the following: If looking at the docker output, there are no further messages. |
Beta Was this translation helpful? Give feedback.
-
|
Sorry for leaving that out. Image: ghcr.io/blakeblackshear/frigate:0.16.0-beta3-tensorrt-jp6 |
Beta Was this translation helpful? Give feedback.
-
|
Interesting, maybe the version of ffmpeg compiled for jp6 doesn't include libwebp. |
Beta Was this translation helpful? Give feedback.
-
|
Looking at the outputs, this looks like to be the case. libwebp is not in the supported encoders, as far as I can see. ffmpeg -versionffmpeg -buildconfffmpeg -encoders |
Beta Was this translation helpful? Give feedback.
-
|
this is fixed in #18740 |
Beta Was this translation helpful? Give feedback.
-
|
Wow, that was fast. Thanks! |
Beta Was this translation helpful? Give feedback.
-
|
Reference: #18738 tldr; my HA alpr entities do not ever go "blank"
|
Beta Was this translation helpful? Give feedback.
-
|
The next issue I'm seeing: There are some parts in the timeline, where the camera feed in the review tab doesn't load the video. Whats puzzling is, currently only one exact hour is affected this time: exactly 9pm to 10pm. If I jump to any other timecode on that day, it works fine. In the nginx log, I have the following error messages - always appearing in pairs. Environment:
|
Beta Was this translation helpful? Give feedback.
and others
Uh oh!
There was an error while loading. Please reload this page.
-
Beta Documentation: https://deploy-preview-16390--frigate-docs.netlify.app/
Images
ghcr.io/blakeblackshear/frigate:0.16.0-beta3ghcr.io/blakeblackshear/frigate:0.16.0-beta3-standard-arm64ghcr.io/blakeblackshear/frigate:0.16.0-beta3-tensorrtghcr.io/blakeblackshear/frigate:0.16.0-beta3-rkghcr.io/blakeblackshear/frigate:0.16.0-beta3-rocmghcr.io/blakeblackshear/frigate:0.16.0-beta3-tensorrt-jp6Changes since Beta 2
Bugfixes
Features
separatorfield to proxy config to specify the character that separates fields in proxy headersMajor Changes for 0.16.0
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.dbfileThe go2rtc audio handling has been updated to accept any audio codec instead of only AAC. This makes it easier to get audio from cameras that use PCMA/U audio. However, if the recording ffmpeg preset
preset-record-generic-audio-copyis currently used it likely will need to be changed topreset-record-generic-audio-aacto continue to support audio in recordings.Object detection is now disabled by default. A config migrator will update your config globally with
enabled: truefordetect:. If you use Frigate with a read-only config, you may need to edit your own config to enable object detection manually.Frigate will no longer fall back to look for a ffmpeg binary in
PATH. If you are extending the Frigate image or somehow injecting a custom ffmpeg binary in a place like/usr/local/bin, please read the custom ffmpeg build guide.Home Assistant Add-on users will need to be aware of several important breaking changes. See the Important Notes for Home Assistant Add-on Users section below.
The base image for Frigate has been updated to Debian Bookworm. The Jetpack 4 (
jp4) and Jetpack 5 (jp5) images have been removed due to not being compatible with the new base image. A new Jetson Jetpack 6 (jp6) build has been contributed by a community member.The camera group icon pack has been updated. Some of your camera group icons may need to be re-selected.
The Hailo driver and firmware has been updated to 1.20.1. The standard Docker image now provides Hailo support. If you are using Home Assistant Operating System, version 15.0 is required. Other users will need to update the Hailo driver and firmware on their system using the user installation script.
The ROCm migraphx detector has been removed: Instead of separately supporting ROCm via MigraphX and via ONNX, the ROCm (
rocm) detector has been removed, and theonnxdetector should be used instead. This change should only require changing thetype: rocmtotype: onnxindetectorsAn additional parameter has been added to
movement_weightsfor autotracking users. A config migrator will add the new value so that Frigate starts, but if this fails or you are usingrelativezooming with autotracking, you should removemovement_weightsfrom your config and recalibrate.Proxy header mapping for user roles is now required or setting the
default_roletoadminis required to achieve the same functionality as previous versions. This version implements a Viewer role, which is now the default for authenticated users without aroleheader map defined in the config. If you are using a proxy, you must includeadminin your mapped role header in order for admin users to retain their elevated privileges. See the documentation.The
strftime_fmtconfig option is now deprecated and will be removed in a future version. The UI will now define time formats based on your language and locale.Some API routes are now protected and require credentials or changes. With the introduction of the Viewer role and the ability to mark review items as reviewed separately per-user, changes may be needed to any services making calls to Frigate's API.
Important Notes for Home Assistant Add-on Users
For Home Assistant add-on users only
The Frigate config directory for add-on users was changed from the Home Assistant config directory to the add-on config directory. With this change, the Frigate config file and database are now part of the add-on backup taken by Home Assistant, no longer needing to backup them manually. Note that only backups after 0.16 will benefit of this change. If you took a backup of Frigate 0.15 when updating to 0.16, the Frigate config file and database will not be part of it. In that case, you should still backup them manually before updating to Frigate 0.16.
Automatic migration
Upon the first Frigate 0.16 initialization, all Frigate-related files will be moved from the Home Assistant config directory to the Frigate add-on config directory. This will mean these files will no longer be accessible through your Home Assistant
/configdirectory. They will now be found in the/addon_configsdirectory, which can also be accessed through the Terminal or VS Code add-ons.In the new directory, the Frigate config file will be named config.yml rather than frigate.yml. This aligns with people running Frigate through other means. If you want to move back to Frigate 0.15 and did not backup the config file, you can copy the
config.ymlfrom the/addons_configback to/config/frigate.ymlbefore restoring the add-on. Consider copying back other files as well, like custom model, custom ffmpeg, and custom go2rtc. Note that the database from 0.16 is not compatible with Frigate 0.15 anyway and thus should not be copied back. To restore the database, you must restore it from a manual backup.Moving the config directory between add-on variants
The
/addon_configsdirectory contains the config directories of all add-ons. You can move files from one addon config directory to another, like from Frigate Beta to the stable one, or vice-versa. Make sure you do that before uninstalling the add-on, unless you do not select the Also permanently delete this addon's data when uninstalling it.Breaking change
If your Frigate config file refers to any arbitrary files in the Home Assistant config dir, like
echo:/config/scripts, their references must be manually updated to from/config/to/homeassistant/.New Features
Frigate 0.16 introduces several major new features.
Face Recognition
Frigate utilizes face detection and recognition to apply sub-labels to known faces. Recognized faces can be included in built-in notifications and integrated into third-party notifications.
Frigate+ is not required to use Face Recognition. However, using a Frigate+ model (or a custom model that detects
facenatively) enhances efficiency by eliminating the need for separate face detection per person.The accuracy of face recognition heavily depends on the quality of the images used for training. It is highly recommended to follow the training guide to achieve the best results.
Refer to the Face Recognition documentation for more.
License Plate Recognition (LPR)
Frigate detects and recognizes license plates, saving recognized plates to tracked car objects. Known license plates can also be assigned custom labels (e.g.,
AB12345-->Bob's Car).Frigate+ is not required for License Plate Recognition. However, using a Frigate+ model (or a custom model that detects
license_platenatively) improves efficiency by eliminating the need for a separate detection model.For additional features such as regex-based matching, refer to the License Plate Recognition documentation.
Viewer Role
Frigate 0.16 introduces a Viewer role that restricts access to system features, the configuration editor, and the Frigate HTTP API. Viewers have permissions to use the Live view, Review, and Explore.
Additionally, each user can mark review items as reviewed independently so that one user's action will not hide the review items from others.
The Users tab in Settings can be used to create users and assign roles, and these roles are enforced when accessing Frigate through the authenticated port. Documentation
More fine-grained permissions (like denying access to specific cameras, for example) will come in a future version.
Proxies can also now map user roles. See the Proxy configuration documentation.
Improved Support For Object Detection Hardware and Models
Hailo8
The Hailo8 detector has been refactored by the Hailo team to greatly improve object detection performance and to add support for more model types. Documentation
ONNX & OpenVINO
Support has been added for a variety of new models for ONNX and OpenVINO detectors
RockChip (RKNN)
RockChip has had the following improvements:
Camera Improvements
Dynamically Enable and Disable Cameras
Frigate now offers the ability to enable / disable cameras dynamically, including buttons in the UI, MQTT, and using the
camera.turn_on/camera.turn_offservices in the Home Assistant Integration. Disabling a camera stops all of Frigate's decoding and processing for that camera. See the documentation.Live Stream Selection
Frigate now allows configuring multiple streams per camera, so that any number of streams for a single camera can be created and selected in the UI. See the Live view documentation.
Camera groups now include per-camera settings to further customize the live streaming behavior of that camera in that camera group.
Other noteworthy live streaming improvements:
Other UI Changes
There are many other notable improvements to the Frigate UI:
Ctrl+ scroll wheel on desktop.i18n Support
Frigate now supports internationalization and localization, making it possible to change the language of the UI. The language is auto-selected based on your browser locale and can also be changed via the Languages menu in the UI.
Some languages are not fully translated. Translation contributions are welcomed and can be made using Weblate.
Backend Features and Performance Improvements
New Base Image
Zones and Object Area
min_areaandmax_areaas a decimal between 0.000001 and .99 to represent a percentage of the frame. Numbers above 1 will still represent areas in pixels. For reference, 0.000001 represents ~1 pixel at 1280x720.Semantic Search
Bird Classification
ONVIF and Autotracking
personobject.Other notable changes
/recordings/summary, to return true/false for each day that recordings are available.This discussion was created from the release 0.16.0 Beta 3.
Beta Was this translation helpful? Give feedback.
All reactions