Vision and Imaging Solutions

HyperAcuity.com

The original Shape

Refined Shape and Detected High Curvature Points

(The original shape is super imposed in gray)

Shape provides one of the most important information in visual processing.  Consequently, it is one of the most common ways to represent objects, and many existing computer vision algorithms rely on good shape representations.  When we develop a new algorithm, we often assume that good representations have been obtained by either low level feature extraction techniques or a human operator. However, the assumption does not always hold and the performance of subsequent tasks can degrade significantly with even a small amount of noise and aliasing present in the  representation.  Thus, it is extremely important to have a general-purpose shape refinement algorithm that is robust against noise and aliasing and preserves salient structures of the shape.

 

Motivation

Example Result

Technical Detail

Download our technical report here.

Demo Program

Download our demo application for Windows here.

With our approach, each contour point is assigned a 2D parametric curve segment, which is used to represent the local portion of the shape between the contour point and its two adjacent ones.  We use a quadratic polynomial to represent a local shape, and call it elastic quadratic wire or EQW. A quadratic polynomial is the least order polynomial with non-constant curvature, thus is able to describe the underlying structure of the shape better than a linear polynomial while less prone to overfitting to noise than a cubic one……  (For more, see our technical report below.

Approach

Basic Research

Shape Representation