Export
Research Overview
Yossi Rubner focuses on computer vision, specifically content-based image retrieval and the mathematical modeling of perceptual similarity. His primary contributions introduced the Earth Mover's Distance to the vision community as a robust metric for comparing color and texture distributions. The most recent papers in the evidence explore empirical evaluations of dissimilarity measures and the application of these metrics to unsegmented image retrieval.
Research Trajectory
Yossi Rubner’s research trajectory in the late 1990s and early 2000s centered on developing and evaluating robust mathematical frameworks for comparing images based on their visual content. A core problem in content-based image retrieval at the time was finding a distance measure that accurately reflected human perceptual similarity, particularly when dealing with complex distributions of color and texture.
In 1998, Rubner led the development of a novel approach to this problem in A Metric for Distributions with Applications to Image Databases. As first author, he introduced the Earth Mover's Distance (EMD) to the computer vision community. The EMD was framed as a special case of the transportation problem from linear optimization, measuring the minimal amount of "work" required to transform one distribution's mass into another. Rubner demonstrated that the EMD was highly effective for image databases because it allowed for partial matching and, when comparing distributions of equal overall mass, functioned as a true metric with easy-to-compute lower bounds.
Building on this foundational metric, Rubner applied the EMD to specific challenges in texture analysis. In the 1999 paper Texture-based Image Retrieval without Segmentation, he argued that traditional image segmentation was not only unnecessary for texture-based retrieval but could actively degrade performance. Instead, he proposed describing images via distributions of spatial frequency descriptors. By utilizing the EMD with an appropriate ground distance, this method successfully handled multi-textured queries, demonstrating the ability to group similar animals together while accurately distinguishing between subtly different coat patterns, such as those of cheetahs and leopards.
Alongside developing new metrics, Rubner participated in rigorous, large-scale benchmarking of existing techniques. As second author on Empirical Evaluation of Dissimilarity Measures for Color and Texture (1999), he contributed to an extensive comparative study of nine image dissimilarity measures. This work, which summarized over 1,000 CPU hours of computational experiments, utilized a novel random sampling scheme for color and an image partitioning method for texture to establish ground truth. The study provided quantitative performance evaluations across classification, retrieval, and unsupervised segmentation tasks, highlighting how the careful selection of a dissimilarity measure drastically improves system performance.
Current Directions
The most recent paper in the provided evidence is from 2000. At that time, Rubner’s research culminated in a comprehensive formalization of his earlier work for the International Journal of Computer Vision. In The Earth Mover's Distance as a Metric for Image Retrieval., he investigated the deeper properties of the EMD for content-based image retrieval. This direction focused on combining the transportation problem solutions from linear optimization with vector quantization representation schemes. By integrating these techniques, the work established an image comparison framework designed to account for perceptual similarity more effectively than previously proposed methods.
Profile updated
Loading new papers…
Loading research neighbors…
See something wrong? Report a mistake