The existing approaches for instance segmentation task can be summarized into three categories:
1、Top-down methods
Paradigm : Solving the problem from the perspective of object detection, i.e., detecting first and then segment the object in the box.
FCIS: Fully convolutional instance-aware semantic segmentation
Mask R-CNN
Path aggregation network for instance segmentation
Mask scoring R-CNN
TensorMask
YOLACT
BlendMask
Mask encoding for single shot instance segmentation
PolarMask
2、Bottom-up methods
Paradigm : Viewing the task as a label-then-cluster problem, e.g., learn the per-pixel embeddings and then cluster them into groups.
Associative embedding End-to-end learning for joint detection and grouping
Semantic instance segmentation with a discriminative loss function
SGN: Sequential grouping networks for instance segmentation
SSAP: Single-shot instance segmentation with affinity pyramid
3、The latest direct method
Paradigm : ({Backbone + FPN + FCN} + [Category Branch + Mask Branch])
Aiming at dealing with instance segmentation directly, without dependence on box detection or embedding learning. （anchor-free , proposal-free and grouping-free）
SOLO: Segmenting objects by locations
SOLOv2: Dynamic, Faster and Stronger
