web: https://sites.google.com/view/mauriciohessflores
LinkedIn: http://www.linkedin.com/in/mauriciohessflores
Welcome to my research website, corresponding to the work I performed as a Postdoctoral Scholar (2012 - 2014) and graduate student (2006 - 2011) at the Institute for Data Analysis and Visualization (IDAV) at UC Davis, with Prof. Ken Joy as advisor. I obtained my PhD in the Department of Electrical and Computer Engineering.
Between 2007 and 2011, I worked with Prof. Joy and Dr. Mark A. Duchaineau at Lawrence Livermore National Laboratory on large-scale scene reconstruction from aerial video. My PhD dissertation is titled 'Error Detection, Factorization and Correction for Multi-View Scene Reconstruction from Aerial Imagery', and was the result of work performed jointly between IDAV and LLNL as part of the Lawrence Scholar Program fellowship at LLNL. My postdoctoral research was a continuation of that work, with an additional component related to the visualization of uncertainty in multi-view scene reconstruction stages. Most recently, we developed algorithms for multi-view triangulation and pose estimation by use of an L1 cost function based on angular error.
I had previously completed an MS in Electrical Engineering at the Universidad de Costa Rica in 2004, working on facial feature extraction and face-model adaptation for model-based coding. My advisor was Dr. Geovanni Martinez, director of the IPCV-LAB. A description of my work and publications there can be found here.
My PhD research involved parameter error detection, classification and correction in multi-view scene reconstruction. Our main focus had initially been on aerial imagery, but our algorithms were designed to work across different real and synthetic scene types, as well as different camera motions. Our initial work dealt with exploring algorithms to help optimize performance, robustness and generality of a novel system that performs video frame-to-frame dense correspondence for applications such as motion analysis, tracking and multi-view imagery, based on feedback after pose and structure estimation to detect and correct errors in this process. Another line of work dealt with improving bundle adjustment conditioning in multi-view stereo by using ray divergence information when attempting to compute scene structure. Given accurate feature matches, divergence is a function of camera parameter inaccuracies. We have shown that this information can be used to weight bundle adjustment such that final reprojection error and processing time can be reduced with respect to other weighting schemes from the literature. We have also designed an algorithm for non-parametric sequential frame decimation, which allows the filtering of frames ill-posed for pose and structure estimation as well as feature matching in long video sequences without using thresholds for the decimation decision. More recent work involved the visualization of uncertainty in multi-view scene reconstruction stages, as well as multi-view triangulation and pose estimation based on an L1 angular error-based cost function. More detailed explanations of each line of work can be found using the links below.