English | 简体中文
In this document, we introduce the protocol used by this toolkit to encode panoptic segmentation labels.
Basically, every single pixel in a panoptic segmentation map (e.g. the reference labels or the model prediction) has the value pan_id. Given the number of classes c and the maximum number of instances n, pan_id ranges from 0 to c * label_divisor + n.
For thing classes, there is:
pan_id = (cat_id + 1) * label_divisor + ins_id
where cat_id is the category ID (starts from 0) and ins_id is the instance ID (starts from 1).
For stuff classes, there is:
pan_id = (cat_id + 1) * label_divisor
For unknown labels, pan_id is equal to 0.
label_divisor is a pre-defined constant. The default value of label_divisor is 1000.