Evaluation · Segmentation
Inspect performance on held-out reference images and find concrete mistakes.
Before you start
A completed Segmentation result and available evaluation images.
Quick start
- Open Evaluation and choose Segmentation in the backend selector.
- Select the run and evaluation data, then run evaluation. Inspect missing objects, extra objects and wrong classes.
- Inspect individual predictions and preserve useful failure evidence for review.
Functions and choices
Keep comparisons fair
Use consistent splits and thresholds when comparing results. Training loss alone does not establish detection or mask quality.
Compare annotations
Human annotations (solid lines) and Model suggestions (dashed lines) use project class colors and can be shown independently to inspect overlapping outlines.
Result
Saved evaluation evidence plus examples that explain the metrics.
Next step
Compare against an earlier model or review mistakes in Failure Bank.