SOLO: A Simple Framework for Instance Segmentation

[JisuHann]

SOLO: A Simple Framework for Instance Segmentation

• task: instance segmentation

  Idea

: the center of an objects falls into a grid cell --> 1) predicting the semantic category 3) segment the object

Semantic category

• predict the C(class num)-dimensional output to indicate the semantic class probabilities

  Instance Mask

• each positive grid cell generate the corresponding instance mask
• using FCN, problem: have to make model more spatially variant
  • Solution: directly feed normalized pixel coordinates to the network (add the spatial functionality, giving the convolution access to its own input coordinates)

    Forming instance segmentation: Category prediction + corresponding mask

• raw instance segmentation results are generated by gathering all grid results
• NMS

Types

• Vanilla SOLO
• Decoupled SOLO
• Dynamic SOLO
• Decoupled Dynamic SOLO

Inference Pipeline

• Semantic category - Category score
• Instance mask - Maskness: quality and confidence of mask prediction