Abstract:Earth Mover's Distance (EMD), targeting at measuring the many-to-many distances, has shown its superiority and been widely applied in computer vision tasks, such as object recognition, hyperspectral image classification and gesture recognition. However, there is still little effort concentrated on optimizing the EMD metric towards better matching performance. To tackle this issue, we propose an EMD metric learning algorithm in this paper. In our method, the objective is to learn a discriminative distance metric for EMD ground distance matrix generation which can better measure the similarity between compared subjects. More specifically, given a group of labeled data from different categories, we first select a subset of training data and then optimize the metric for ground distance matrix generation. Here, both the EMD metric and the EMD flow-network are alternatively optimized until a steady EMD value can be achieved. This method is able to generate a discriminative ground distance matrix which can further improve the EMD distance measurement. We then apply our EMD metric learning method on two tasks, i.e. , multi-view object classification and document classification. The experimental results have shown better performance of our proposed EMD metric learning method compared with the traditional EMD method and the state-of-the-art methods. It is noted that the proposed EMD metric learning method can be also used in other applications.

Shape context based on H-EMD algorithm for shape matching

3D Shape Matching Using Collinearity Constraint

Shape Representation and Recognition in High Dimensional Feature Space

Radius-normal Histogram and Hybrid Strategy for 3D Shape Retrieval

Shape matching and object recognition using shape contexts

Gesture Recognition Based On Improved Shape Context Algorithm And Earth Mover'S Distance

A Progressive Approach for Computing the Earth Mover's Distance

Run-length based shape descriptor and shape matching technique based on the HMM Model

The Generalized Shape Distributions for Shape Matching and Analysis

Contour Guided Hierarchical Model For Shape Matching

Shape Distribution Algorithm Based on Geodesic Distance

Embedded Profile Hidden Markov Models for Shape Analysis

A New Shape Signature for 3D Model Similarity Assessment

Halcon Application for Shape-Based Matching

Articulated 3D model matching using multi-scale histograms of shape features for customized additive manufacturing

Contour Based Shape Matching For Object Recognition

Fast Shape Matching And Retrieval With Dynamic Angular Partition Descriptor

EMD Metric Learning

Shape matching and retrieval based on multiple feature descriptors

Shape Point Set Matching Based on Oriented Shape Context in Turbulence-Cluttered Scene

A robust CPD approach based on shape context