Open imneonizer opened 2 months ago
- Does it output instance segmentation masks?
- Can we use the trained weights with automatic mask generator?
My goal is to do field delineation, I already got success with default weights using computer vision post processing, but I wanted to improve it further, training sam on domain data seems like a perfect solution to me. Only issue is it segments everything including roads water bodies etc, I want it to focus only on the fields.
Single produced mask on a 1024x1024 window:
Produced masks interpolated over a large area:
Converted to boundaries:
This can be extended to instance segmentation. We use this for water segmentation from the background, and it can be. Please try to use this fine-tuning for more than one class. The results you provided are already good, as well.
My goal is to do field delineation, I already got success with default weights using computer vision post processing, but I wanted to improve it further, training sam on domain data seems like a perfect solution to me. Only issue is it segments everything including roads water bodies etc, I want it to focus only on the fields.
Single produced mask on a 1024x1024 window:
Produced masks interpolated over a large area:
Converted to boundaries: