Reproduce “Rotation equivariant vector field networks” on OpenResearch

About this paper

Diego Marcos, Michele Volpi, Nikos Komodakis, Devis Tuia

University of Zurich researchers developed Rotation Equivariant Vector Field Networks (RotEqNet), a CNN architecture that explicitly encodes rotation equivariance, invariance, and covariance. The method efficiently processes rotated features by transforming filter responses into compact 2D vector fields, allowing it to achieve competitive or superior performance across tasks like image classification, segmentation, and orientation estimation with significantly fewer parameters than existing models.

Key insights

• Propagating orientation information as a 2D vector field (magnitude and angle) through an 'orientation pooling' mechanism drastically reduces feature map dimensionality while preserving critical rotational information.
• The architecture enables the construction of deep networks that explicitly handle rotation equivariance, invariance, and covariance, moving beyond implicit learning via data augmentation.
• Learning a single canonical filter and rotating it on-the-fly, combined with gradient aggregation, allows for significantly fewer learnable parameters compared to standard CNNs or other rotation-equivariant methods.

Summary by alphaXiv.