Harmonic Filters for Generic Feature Detection in 3D

Abstract

This paper proposes a concept for SE(3)-equivariant non-linear filters for multiple purposes, especially in the context of feature and object detection. The idea of the approach is to compute local descriptors as projections onto a local harmonic basis. These descriptors are mapped in a non-linear way onto new local harmonic representations, which then contribute to the filter output in a linear way. This approach may be interpreted as a kind of voting procedure in the spirit of the generalized Hough transform, where the local harmonic representations are interpreted as a voting function. On the other hand, the filter has similarities with classical low-level feature detectors (like corner/blob/line detectors), just extended to the generic feature/object detection problem. The proposed approach fills the gap between low-level feature detectors and high-level object detection systems based on the generalized Hough transform. We will apply the proposed filter to a feature detection task on confocal microscopical images of airborne pollen and compare the results to a 3D-extension of a popular GHT-based approach and to a classification per voxel solution.