Shape Similarity Measurement for Known-Object Localization: A New Normalized Assessment.

Baptiste Magnier,Behrang Moradi

doi:10.3390/jimaging5100077

Abstract

This paper presents a new, normalized measure for assessing a contour-based object pose. Regarding binary images, the algorithm enables supervised assessment of known-object recognition and localization. A performance measure is computed to quantify differences between a reference edge map and a candidate image. Normalization is appropriate for interpreting the result of the pose assessment. Furthermore, the new measure is well motivated by highlighting the limitations of existing metrics to the main shape variations (translation, rotation, and scaling), by showing how the proposed measure is more robust to them. Indeed, this measure can determine to what extent an object shape differs from a desired position. In comparison with 6 other approaches, experiments performed on real images at different sizes/scales demonstrate the suitability of the new method for object-pose or shape-matching estimation.

Highlights

Introduction and MotivationsRepresenting an object shape is extremely useful for specific industrial and medical inspection tasks
This edge detection evaluation measure separately computes the distances of False Positive points (FPs) and False Negative points (FNs) in function of the number of points in detected contour map (Dc) and ground truth (Gt), respectively, but it is not normalized; so its scores are interpretable with difficulty (Appendix A of this paper presents other non-normalized measures with results regarding real videos Video 2 (V2), Video 3 (V3) and video 4 (V4).)
A new approach to measuring a contour-based object pose is presented in this paper

Summary

Introduction

Representing an object shape is extremely useful for specific industrial and medical inspection tasks. When a shape is aligned, under supervision, with a reference model, a wide variety of manipulations can arise/be used. Contrary to region-based methods [1], edge-based representation remains a set of methods only exploiting information about shape boundaries. The assessment of acquired features (contours) in a candidate image compared to an ideal contour map model is one approach to the supervised assessment of shape depiction. This paper presents a new approach for the measurement of a contour-based object pose, which is normalized. It follows on from a talk given by the research team in [2], dealing with the subject more thoroughly and in greater detail

Objectives

Methods

Results

Conclusion