Coffee Cup Detector
Collaborators: Galen Brown, Nicholas Lanotte, Fan Mo
In this report, a classification algorithm for detecting coffee cups in an image is shown for the purposes of demonstrating computer vision techniques as part of the requirements for the CS/REB 549 Computer Vision course at Worcester Polytechnic Institute. Images of coffee cups are further classified by the material the coffee cups are made from using a second classification algorithm to add complexity to the project. Due to the high degree of variation in color and texture, and a moderate degree of variation in the shape of the target objects, we selected feature-based Support Vector Machines and adversarially trained Convolutional Neural Networks to perform the initial classification. A Decision Tree was then used to further classify images identified as containing a coffee cup. The binary classification algorithms were able to perform with above 80% accuracy and the Decision Tree had a similar performance of 78% accuracy when used by itself. However, the Decision Tree was unable to generalize enough to ignore complex backgrounds which made its performance poor when classifying images identified as containing a coffee cup from the binary classifiers. We discuss our methodology, results, conclusions, and future directions for the project.
Real Time Tennis Match Prediction using Machine Learning
Collaborators: Gabriel Entov, Nicholas Lanotte, Adam Santos
This paper discusses a supervised learning algorithm trained to predict tennis matches by combining both historical match data and real-time statistics. Different models for predicting matches before they begin were explored and compared to each other. A separate algorithm for predicting a match as it is being played was created by testing and comparing multiple algorithms and data structures. With both of these algorithms established, they were combined to produce an algorithm with better performance overall than either of the two algorithms alone. Sets were found to be the most dominant and predictive feature for matches, with other standout features also revealed, such as service points, specifically on the first serve. Finally, the project outlines the future work needed to help combat the problems of the classifier such as predicting upsets and comebacks.
Collaborators: Jake Kelley, Dong Hyun Ko, Samantha E. Robinson
Inconsistent and excessive dwell times often cause delays to underground metro systems. The goal of this project was to recommend solutions to minimize dwell times in the Central Line of the London Underground. The team conducted employee interviews, station observation, train observation, CCTV observation, and passenger surveys to more thoroughly understand the issue. Our team identified four system constraints that could be altered to encourage more efficient passenger behavior in order to reduce dwell times.
Shifting Reality Template
A Unity project template that contains the building blocks for creations that shift between AR and VR on the Quest family of headsets by utilizing their passthrough video capabilities.