US tab

Ruben Villegas

Undergraduate Majors:
Computer Science
Graduate School Plans:
Machine Learning

Ruben Villegas

Ruben Villegas was born in Machala, Ecuador. He is majoring in Computer Science with a minor in Mathematics. Ruben has participated in an NSF REU in Computer Vision at UCF as well as in the University of Michigan Summer Research Opportunity program (SROP) in Machine Learning. He currently works in the Center for Research in Computer Vision at UCF as an undergraduate researcher under the supervision of Dr. Mubarak Shah. Ruben is also a Google/Hispanic Scholarship Fund scholar. His goal is to attain a PhD in the field of Machine Learning, mainly Deep Learning. He plans to apply this knowledge to the Computer Vision area, more specifically, Human Face Recognition.

Title: Convolutional Neural Networks for Vehicle Detection

Conducted at the University of Michigan, Ann Arbor, as part of their Summer Research Opportunity Program

Mentor: Dr. Honglak Lee, University of Michigan - Ann Arbor

Co-authors: Hongji Wang, and Charlie Yan

Car detection in still images is a very important problem in the Computer Vision field that is needed in applications ranging from driver assistance systems to autonomous cars. The Perceptual Robotics Laboratory (PeRL) at the University of Michigan provided the dataset used to train and test our method, for which we used Hierarchical Clustering to extract and separate positive and negative examples to be used in our training. Our method uses a combination of static image car detection reinforced with point cloud information and a trained Convolutional Neural Network (CNN) for car detection. For our static image classifiers we extract Histogram of Oriented Gradients (HOG) features, while for our point cloud classifiers we extract Spin Image (SI) features. We feed each of our extracted features to a Support Vector Machine (SVM) to generate our classification models. We also train a CNN for detection and a Radial Basis Function (RBF) using either 2D and CNN or 3D and CNN detection features, positive and negative. The way our pipeline is set up is, we first compute the detections using a CNN, followed by computing the 2D detections using HOG and the 3D detections using SI on the CNN detections to narrow down our detections. Finally we train our RBF model using the 3D and CNN detection features on the previously detected cars, and evaluate our pipeline using the learned RBF model.

Title: Human Detection Using Spatio-Temporal Pictorial Structures

Conducted at the University of Central Florida as part of the National Science Foundation Research Experience for Undergraduates (REU) program

Mentors: Dr. Mubarak Shah, and Dr. Niels Da Vitoria Lobo, University of Central Florida

Co-authors: Amir Roshan, and Afshin Dehghan

Human detection is a very important and challenging problem in computer vision, which has many applications. This paper describes a method to detect humans in videos extending a frame-by-frame detection to take in consideration transitions of human parts in time. The experiments performed on the TRECVID MED11 dataset using preprocessed detections in each frame, show our proof of concept that including the temporal information will improve human detection in videos. Learning a 3d-model that relates human parts in space and though time, allows us to limit the number of false detections in space by adding the constrain of smooth transitions through time.