Few-example affine invariant ear detection in the wild

Abstract

Ear detection in the wild with the varying pose, lighting, and complex background is a challenging unsolved problem. In this paper, we study affine invariant ear detection in the wild using only a small number of ear example images and formulate the problem of affine invariant ear detection as a task of locating an affine transformation of an ear model in an image. Ear shapes are represented by line segments, which incorporate structural information of line orientation and line-point association. Then a novel fast line based Hausdorff distance (FLHD) is developed to match two sets of line segments. Compared to existing line segment Hausdorff distance, FLHD is one order of magnitude faster with similar discriminative power. As there are a large number of transformations to consider, an efficient global search using branch-and-bound scheme is presented to locate the ear. This makes our algorithm be able to handle arbitrary 2D affine transformations. Experimental results on real-world images that were acquired in the wild and Point Head Pose database show the effectiveness and robustness of the proposed method.