Evaluating Segmentation Models

Evaluating Segmentation Models

Judge a mask/alpha model on metrics that reflect perceived quality, not a single average that hides edge failures. Covers hard segmentation and soft matting, and the parity check that a converted or quantized model still passes.

When to use

• Sign-off on a trained segmentation / matting checkpoint.
• A/B two checkpoints, or before-vs-after a conversion/quantization step.
• Deciding whether a model is "good enough" to export or ship.

Metrics (pick by task — see references/metrics.md for formulas)

• Hard segmentation: per-class IoU + mIoU; pixel accuracy (report but don't rely on it under class imbalance); boundary-F (F-score of boundary pixels within a tolerance) — the metric that actually tracks edge quality.
• Image matting (soft alpha): the four standard metrics SAD, MSE, Grad, Conn, computed in the trimap unknown band (the contested region) as well as whole-image; plus binary IoU of the thresholded alpha for a coarse sanity check.

How to evaluate

1. Per-class / per-region, not one number. A single mIoU hides a collapsed rare class; a whole-image SAD hides bad hair/edges. Always break down by class and by the boundary/unknown band.
2. Edge-vs-global diagnostic. If mean error is low but max_abs is near 1.0 at sparse edge/thin-structure pixels, the model is globally close but locally wrong — a classic colour-order or missed-stem symptom (see converting-pytorch-to-tflite), not a real pass.
3. Parity of the metric across host↔device. When evaluating a converted/quantized model, judge consistency in decision space (thresholded mask / argmax / IoU), not raw logits or raw alpha — reduced precision moves logits harmlessly but must not move the decision. (Hand device sign-off to validating-on-device-inference.)
4. Sample selection for the report. Collect all failure cases; sample the correct ones. Feed the numbers + galleries to rendering-html-eval-reports — this skill decides what to measure, that one renders it.

Acceptance

• State an explicit threshold per metric (e.g. mIoU ≥ X, boundary-F ≥ Y, matting SAD ≤ Z in the unknown band) before looking at results; a metric without a pre-set bar is a description, not a gate.
• A pass needs the per-class/boundary breakdown to clear the bar — not just the average.

Boundaries

• Read/measure only: emits PASS/FAIL rows + numbers; it does not train or tune.
• Class set, thresholds, trimap source, and dataset are caller inputs — keep zero project values in any eval you run. Rendering → rendering-html-eval-reports; on-device sign-off → validating-on-device-inference.