Idea

Inherited from BiSeNet v1:

  • Detail Branch, with wide channels and shallow layers to capture low-level details and generate high-resolution feature representation
  • Semantic Branch, with narrow channels and deep layers to obtain high-level semantic context (lightweight with a fast-downsampling strategy)

Improvements

  • Guided Aggregation Layer to enhance mutual connections and fuse both types of feature representation
  • Booster Training Strategy

Architecture

Context Embedding Block

Uses the global average pooling and residual connection to embed the global contextual information efficiently

Guided Aggregation Layer

Booster Training Strategy

Insert the auxiliary segmentation head to different position of the Semantic Branch, enhancing the feature representation in the training phase