Panoptic Segmentation
lastmod 2024-02-06created 2020-02-142 edits
Roughly Semantic segmentation + Instance segmentation.
About
Panoptic segmentation can be seen as combining semantic segmentation (labeling each pixel) and instance segmentation (recognizing and segmenting objects). In other words, panoptic segmentation is a concept that also includes distinguishing stuff from things.
The 'Panoptic Segmentation' paper also proposes metrics for this task. To understand panoptic segmentation, you need to understand the concepts of stuff and things used in the paper. Countable objects (countable objects) distinguished by instance segmentation are called things, and amorphous, uncountable regions (amorphous and uncountable regions) handled through semantic segmentation are called stuff. The existing metrics used for semantic segmentation and instance segmentation cannot be used for panoptic segmentation, because they do not consider stuff and things classes at the same time. To help understand the metrics proposed in the paper, let's look at IoU (Intersection of Union) and one of the paper's figures.
Documentation
- Panoptic Segmentation
- https://arxiv.org/abs/1801.00868