Capsule Pooling is a method for aggregating capsule outputs in capsule networks that aims to preserve the spatial and hierarchical structure of features.
Unlike conventional pooling methods such as max or average pooling, which often simplify information significantly, capsule pooling combines vectors or matrices that contain pose and probability information.
By pooling at the capsule level, models can reduce complexity while preserving important relationships between objects or parts.
Typical approaches include weighted merging of capsule outputs based on activation strengths or routing weights, thereby improving model capacity and generalization.