80 likes | 247 Views
Adam Yeh. UCF Computer Vision REU Week 10. This Week. Improved binary results using mask Expanded database Search photo-hosting sites for expressions Flickr, photobucket, webshots Reprocessed images Zoomed in more on face to eliminate background junk
E N D
Adam Yeh UCF Computer Vision REU Week 10
This Week • Improved binary results using mask • Expanded database • Search photo-hosting sites for expressions • Flickr, photobucket, webshots • Reprocessed images • Zoomed in more on face to eliminate background junk • Area of face in dataset more in tune with output from existing face detection systems • Multi-class classification with NSV-based model
Mask Algorithm • For each class, find variances for each feature (pixel) • Find threshold, say 60th percentile variance value • For feature x, if variance(x)>threshold, mask[x]=1, otherwise mask[x]=0 • Remove the connected components of value 1 that are on borders • Union the results among all classes • Addition to expand number of features looked at
Results • Mask improvements: • Happy vs Surprise: 88% to 89% • Happy vs Neutral: 81% to 90% • Surprise vs Neutral: 55% to 70% • Feature reduction: 576 (24x24) to 218 features
Multi-class classification • Find intersection of NSVs from binary SVMs • 22/66 neutral, 389/444 happy, 79/126 surprise NSVS • Nearest Neighbors • 46% testing accuracy on NSVS • 44% on all data • Happy testing accuracy is 100% while other classes do significantly worse • Large happy dataset? • Distance to Poles (average of NSVs) • 21% testing accuracy on NSVs • Happy testing accuracy is 0%--opposite result of nearest neighbors • Coresvm built-in multi-class classification • One vs. One • 70% testing accuracy without mask • 78% testing accuracy with mask • One vs. All • 71% testing accuracy without mask • 74% testing accuracy with mask
Next Week • Previously: “closeness” measure to poles very simplistic • Simple Euclidean distance • Mahalanobis distance • Better idea of closeness to center of mass • Takes correlation between features in feature vector into account • Scale-invariant