Question about ontology for sam2 #335
-
Search before asking
Notebook nameHi this is more of a question about the ontology can it work to be trained to segment the rear of a car by prompting it "rear of car" for example or is it limited to more simpler learning like a car on it's own but not specific, thanks very much BugNo bug just question EnvironmentWindows 11 Minimal Reproducible ExampleNo response AdditionalNo response Are you willing to submit a PR?
|
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 2 replies
-
|
Moving this to a discussion as it's not a bug. |
Beta Was this translation helpful? Give feedback.
-
|
From my experience - yes, it ought to be able to do that, much in the same way as it can detect faces and not the whole person. Place a positive marker on the car trunk, and multiple negative markers on other car parts. Note that it's focused heavily on specific previously seen objects and may learn to track the back of a specific car. Instead, I'd check out the dataset used in this notebook and train a detection or segmentation model. |
Beta Was this translation helpful? Give feedback.
-
|
Oh wow already exists perfect ! |
Beta Was this translation helpful? Give feedback.
-
|
It can be hard to find, but we have the wildest datasets on Universe 😆 |
Beta Was this translation helpful? Give feedback.
From my experience - yes, it ought to be able to do that, much in the same way as it can detect faces and not the whole person. Place a positive marker on the car trunk, and multiple negative markers on other car parts.
Note that it's focused heavily on specific previously seen objects and may learn to track the back of a specific car.
Instead, I'd check out the dataset used in this notebook and train a detection or segmentation model.