Abstract
Patch descriptors are used for a variety of tasks ranging from finding corresponding points across images, to describing ob ject category parts. In this paper, we propose an image patch descriptor based on edge position, orientation and local linear length. Unlike previous works using histograms of gradients, our descriptor does not encode relative gradi- ent magnitudes. Our approach locally normalizes the patch gradients to remove relative gradient information, followed by orientation dependent binning. Finally, the edge histogram is binarized to encode edge loca- tions, orientations and lengths. Two additional extensions are proposed for fast PCA dimensionality reduction, and a min-hash approach for fast patch retrieval. Our algorithm produces state-of-the-art results on pre- viously published ob ject instance patch data sets, as well as a new patch data set modeling intra-category appearance variations.