Table of Contents
A Radial Contrast Similarity Transform for Robust
Image Correspondence
Motivation
The Problem
The Problem
The Problem
Foreground without contrast
Simple robust norm misses contrast
Edges are ambiguous
Considering contrast
A New Robust Local Transform
A New Robust Local Transform
A New Robust Local Transform
Computing the RCS transform
Radial Cumulative Similarity
Model occlusion but not background
Stable despite low foreground contrast
RCS transform distance
Finger tracking: window A
Finger tracking: window B
Finger tracking: windows C & D
RCS transform of finger windows
Comparative Results
Comparative Results
Comparative Results
Tracked features frame 0
Ground truth frame 1
L2 norm results
Robust norm results
RCS results
Mouth open, no teeth
Mouth open, teeth showing
L2 norm results
Robust norm results
RCS results
Finger tracking, frame 0
L2 & Robust results
RCS results
Quantitative tracking results
Limitations of RCS
RCS+L2 results
Spatial smoothing with RCS
Conclusions |