- Project Type: Instance Segmentation
- Operating System & Browser: Windows
- Project Universe Link or Workspace/Project ID:
I’m currently working on an Instance Segmentation project involving clusters of caterpillars, and I’m looking for guidance on implementing a soft label system rather than using standard hard labels.
I want the model to express varying confidence levels in its predictions, and have those reflected visually in the “display shape” visualization,not just through filtering by a global confidence threshold. For example, if the model is unsure whether a specific region of a caterpillar is the head or abdomen, I’d like that uncertainty value to be visualized through a gradient or varying opacity, rather than displaying solid, confidence labels across the image I am trying to annotate.
Essentially, Is there a way to have the model output and visualize per-annotation confidence maybe as a part of a heatmap or intensity mask? So that ambiguous regions are clearly distinguished from confident ones? If not, are there any workarounds or integrations (like with Label Studio or custom pipelines) that could help simulate this behavior?
Thank you in advance.