Fourier-Ring Descriptors and Measuring Rotational Equivariance
Tegan Emerson (PNNL)
This talk will introduce Fourier-Ring Descriptors (FRDs): an image feature vector based on Fourier analysis. FRDs are rotationally invariant by construction and address a perennial challenge in machine learning for microscopy and overhead images. I will present novel metrics for quantifying equivariance and compare FRDs to learned featurizations from neural network architectures. These featurization approaches are compared and evaluated on the xView image dataset.