[Detector Support]: Frigate+ model falsely detects cars in many occasions #15383
Replies: 3 comments 24 replies
-
|
you should follow the docs for the steps to do after getting your Frigate+ model https://docs.frigate.video/plus/first_model#step-4-adjust-your-object-filters-for-higher-scores |
Beta Was this translation helpful? Give feedback.
-
|
Alright; I'll give it a try. Thanks. |
Beta Was this translation helpful? Give feedback.
-
|
Here are some examples I wanted to attach in my original post. Just for consistency reasons: |
Beta Was this translation helpful? Give feedback.
-
|
You'd have to ask @blakeblackshear |
Beta Was this translation helpful? Give feedback.
-
|
Once your available trainings hits 0, you will have the option to purchase more. |
Beta Was this translation helpful? Give feedback.
-
|
@mike2307 when I started with Frigate six months I just added the basic filters and thresholds at the start of my config as values to apply globally, and my results, when I later submitted my images and integrated the subsequent model, were very good: After my first model submission I experienced a big improvement, and then again with the next. I now get detection accuracies in excess of (best guess, I expect there's a way to work it out accurately, but not sure how, at the moment) 95%. After the first or second or third training, I modified some of the values slightly, and the above is what I am running now. Later on I added some specific tracked items per camera, for example: or Some of this config is for my "experimental" 0.15 docker compose installation which I only run from time to time during the week, and some is from my original 0.13 which is still my main service running 24/7 . |
Beta Was this translation helpful? Give feedback.
-
|
I would find it very helpful to see a page with the last N images tagged with a specific label and the confidence stats. Like the last 20 cars. That would make it easier for me to see what to set min_score to. Or is there functionality for this that I'm overlooking? |
Beta Was this translation helpful? Give feedback.
-
|
you can use the explore page for this |
Beta Was this translation helpful? Give feedback.
-
|
In my experience so far, when it sees an object it isn't familiar with, it will suggest whatever the most labeled training image is in the set. At first everything was a car but as I have added hundreds of deer labels, now everything is a deer. I'm starting to submit images now without any labels in order to balance out things. |
Beta Was this translation helpful? Give feedback.
-
|
That's not my experience, car is by far the most labeled and I don't have any random objects being detected as a car |
Beta Was this translation helpful? Give feedback.
-
|
I wouldn't say it is completely random, the house across the street used to get labeled as a car all the time until I started putting in deer labels. Now it is getting labeled as a deer. If you try hard you can understand why it looks like a deer with the trees representing legs and house body, etc. I'm not exactly sure how many hidden layers are in the model and the counts, but it seems to be biased towards finding something rather than nothing. I would rather that than the opposite though so it's ok. I've been submitting the scene images now in various daylight/night settings in hopes that it gets used to having nothing to see as well.
|
Beta Was this translation helpful? Give feedback.
-
|
Did you follow https://docs.frigate.video/plus/first_model#step-4-adjust-your-object-filters-for-higher-scores |
Beta Was this translation helpful? Give feedback.
-
|
Yep, I'm constantly tweaking them as well. I just notice the more images I put into the training set of each label the more confident it becomes, however, that also increases the false positive confidence as well. I used to work in the 90s on AI image classification , and we noticed this as well when trying to detect bad welds on assembly line. When a model is trained to identify exceptions, it tends to find more exceptions (bad and good) as the training set grows. As Blake mentioned recently in the other comment, it is quite helpful to give it some diversity of labels. |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Describe the problem you are having
Hi,
first of all thank you for the great work. I'm a frigate user for almost 2 years now and it's great!
Since about 1.5 weeks I'm also a subscriber to frigate+. But here I encountered quite some issues...
At one of my two locations the cameras are located outside in the wild with a surrounding forest. I was really looking forward for deer, rabbit, squirrel, etc. detections.
But what I have at the moment can be best described as some kind of "car mania". The model sees cars in almost anything (like tree branches, cabins, ...).
In total I submitted 781 images in this short time frame (yes, I've spend a LOT of time on it). Out of these images there are 422 (!) false positives on cars!
In between my submissions I requested new models. Now I'm at 4 and it's only sightly getting better. I feel like I'll be ending up consuming all of my 12 requests/year in just a few weeks.
Who can help me out here?
I would like to add more annotated images but I feel like I'll be running out of model requests soon.
I could also share my registered mail address for the account if that helps (it's different to the one on github).
Thanks a lot in advance.
(this is a copy of my original post here #14762 (comment))
Version
0.13.2-6476f8a
Frigate config file
docker-compose file or Docker CLI command
Relevant Frigate log output
Install method
Docker Compose
Object Detector
Coral
Screenshots of the Frigate UI's System metrics pages
Any other information that may be helpful
No response
Beta Was this translation helpful? Give feedback.
All reactions