Odd Leaf Out Combining Human and Computer Vision Arijit Biswas, Computer Science and Darcy Lewis, iSchool Derek Hansen, Jenny Preece, Dana Rotman-University of Maryland’s iSchool David Jacobs, Eric Stevens-University of Maryland Computer Science Jen Hammock, Cynthia Parr-The Smithsonian Institution Refining Metadata Associated with Images Existing Image Crowdsourcing Games How our game is different • Anyone can play and can provide us with useful information. • No expertise necessary • Capitalizes on strengths of humans and algorithms – Humans are better than algorithms at identifying similarity of images Game Mechanics Game Mechanics How Leaf Sets Are Constructed • Designed to bring in useful data • Not too easy or too hard • Curvature based histograms used to get features from leaf shapes. – These features are used to find distance between all possible pairs of leaves. What’s in it for us if people play this game? • Identify errors in the dataset • Discover if color helps humans identify leaves • Feedback on how enjoyable or difficult the game is Game Variations Feedback Mechanism Mul tipl e Gues s es Ski p Contes t a fter Ga me i s Fi ni s hed Contes t Previ ous Round Before Leaf is Chosen When Feedback Occurs After Leaf is Chosen Mechanical Turk Trial 30 Number Correct 25 20 15 10 5 0 1 2 3 Enjoyment 4 5 Mechanical Turk Trial 5 Enjoyment 4 3 2 1 1 2 3 Difficulty 4 5 Summary • Anyone can help in Computer Vision research work. • Games can be fun for players and useful for researchers. • Humans are better than machines in judging the similarity of two images. Funding This work is made possible by National Science Foundation grant number 0968546