Replies: 2 comments
-
|
Which model type are you using? I'm working on some tests to try a few things before the next base model. Also, send me an email if you haven't already so I can run some tests specifically using your images. |
Beta Was this translation helpful? Give feedback.
0 replies
-
|
Also, training a model without the base weights will take 50x the amount of GPU time, and that translates to 50x the cost. I don't think that's practical. |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
With close to 10k personal annotations, I'm still seeing obvious false positives that I can only guess is due to the wide variety of base model annotations in environments / configurations that are not similar to mine and the limited size of the live model. I know I can build my own models from scratch but would rather take advantage of the Frigate+ organization of my images. Or maybe add a parameter into the model on a per camera basis along with the other seemingly inert parameters? If the model was using the camera attributes on the Frigate+ page maybe the "indoor" environment would significantly reduce the likelyhood of racoon identifications inside.
For example I see a lot of Black/White IR images showing up as racoon, when there is a very obvious deer / cats in the frame. Im guessing there is an overwhelming amount of IR racoon images in the base model where the model is leaning towards IR == Racoon. Any intense bright spot is a car, etc. ( I purposefully avoid lights directly in camera training)
The suggestion model for the most part gets it right, so I am guessing the live model is not big enough to handle the variations. No matter how many times I verify suggestions (2,100 postives, 910 false p) the live model never really gets better at it.



The bigger suggestion model gets it right about 95% of the time.
Beta Was this translation helpful? Give feedback.
All reactions