New Framework for Part-Aware Instance Matching in Panoptic Segmentation
A new arXiv paper (2605.31094) redefines instance matching for panoptic segmentation evaluation by framing it as a constrained bipartite assignment problem. The authors systematically explore matching strategies when IoU threshold is below 0.5, identifying four strategies: One-to-One, Many-to-One, One-to-Many, and Many-to-Many. They show the first three are compatible with the Panoptic Quality (PQ) metric, while Many-to-Many is not. This work addresses challenges like fragmented instances, adjacent object delineation, and noisy annotations. The framework introduces vertex-based accounting for true positives, false negatives, and false positives.
Key facts
- Paper arXiv:2605.31094
- Panoptic Quality (PQ) metric is standard for joint instance and semantic segmentation evaluation
- Original PQ relies on One-to-One matching above IoU 0.5
- Below 0.5, multiple matching strategies emerge
- Matching recast as constrained bipartite assignment problem
- Four strategies: One-to-One, Many-to-One, One-to-Many, Many-to-Many
- First three are well-defined within PQ framework
- Many-to-Many falls outside PQ framework
Entities
Institutions
- arXiv