• Calendar

February 12, Tuesday
12:00 – 13:00

Histogram distance functions are the cornerstone of numerous computer vision and machine learning tasks (e.g. image retrieval, descriptor matching and k-nearest neighbor classification). It is common practice to use distances such as the Euclidean and Manhattan norms to compare histograms. This practice assumes that the histogram domains are aligned. However, this assumption is violated through quantization, shape deformation, light changes, etc. The Earth Mover’s Distance (EMD) is a cross-bin distance that addresses this alignment problem. We present several new Earth Mover's Distance variants that are robust to outlier noise and global deformations. Additionally, we present efficient algorithms for their computation. We show state-of-the-art results for descriptor matching and image retrieval. These tools have already been used by other groups and demonstrated state-of-the-art results for a range of tasks such as superpixel matching, descriptor matching, image retargeting, image segmentation, social graph comparisons and population density comparison. Finally, we describe several future directions including learning with general (not necessarily positive-semidefinite) similarity functions.