Justin Uang Rose Hills
Textured-Object Detection and Robotic Perception
Perception, or the ability to perceive the world, is one of the remaining missing components for allowing robots to enter our daily environments. However, its current state-of-the-art performance is far from accurate. This summer, I will explore several machine learning techniques for improving perception. For instance, I will use distance metric learning to improve upon the euclidean distance that is widely used for comparing data points. Preliminary results show that algorithms such as Large Margin Nearest Neighbor can greatly improve k-nearest-neighbor performance for SIFT features, improving perception performance. I also plan to use structured data with unsupervised feature learning to improve the feature descriptors used in perception systems. Namely, the features that we hope to learn will capture local depth structure, albedo, and lighting.