New MMA metric fixes instance segmentation evaluation flaws with global matching
Say goodbye to unreliable scoring – MMA delivers stable, sensitive, interpretable results.
Deep Dive
Researchers propose Maximum Matching Accuracy (MMA), a new metric for instance segmentation evaluation. MMA replaces flawed existing metrics (AP@50, PQ, SEG, AJI) that suffer from hard IoU thresholds, per-object normalization, and greedy matching. It uses globally optimal one-to-one matching and per-pixel normalization, producing scores that are more stable, sensitive, and interpretable. Tested on synthetic failure cases, progressive corruption tests, and a model ranking comparison for biological cell imaging.
Key Points
- MMA replaces hard IoU thresholds with continuous, threshold-free scoring for higher sensitivity.
- Globally optimal one-to-one matching eliminates order-dependent, greedy assignment errors.
- Outperforms AP@50, PQ, SEG, and AJI in stability across split cells, merged cells, and boundary imprecision tests.
Why It Matters
Better evaluation metrics mean fairer model comparisons and faster progress in medical image analysis.