Homepage of Raphael Sznitman

Me
E-mailsznitman@jhu.edu
Phone+ 1 410 516 8604
AddressDepartment of Computer Science
The Johns Hopkins University
CSEB Room 136
3400 North Charles Street
Baltimore, MD 21218, USA

Education:

Research

I am a PhD student in the Computer Science department at Johns Hopkins University (Baltimore, USA). I am part of the LCSR and ERC labs. My academic advisor is Dr. Gregory Hager.

I am mainly interested in computational vision, dealing with visual tracking and object detection. More specifically, I am interested in the use of probabilistic and statistical learning in computer vision contexts.

Here is a short but growing summary of my work:

Active Backgrounds

Over the last two decades, background modeling techniques have focused on representing the general appearance of a background that is assumed to be predominantly static. However, there are many situations in which there are active, moving elements that are effectively part of the background. Examples include tools in manipulative tasks or work settings where a small, fixed set of people are moving about. Such situations are not well modeled by traditional methods. In order to deal with these Active Backgrounds, we have developed a modeling approach, Actors on a Stage (AOS), that is able to accommodate both passive and active backgrounds. AOS is presented as a general, recursive estimation scheme for a background model. In this model, actors are a latent variable that is used to explain both occlusion of, and abrupt changes to a background model.

Currently working on this with Henry Lin and Dr. Gregory Hager

Face Detection and Localization by Playing 20 Questions

In recent years, object detection and localization has been very popular by advances in Branch-and-Bound methods. However, a major problem with the methods proposed so far is their need for a robust feature, such as SIFT or SURF, in order to combine the detection (e.g. Bag of Words detection) and search problem into one. In the context of face detection, finding such robust features is difficult and these methods typically do not make use of excellent face detectors developed in the last decade ( e.g. Viola Jones Face detection). For this reason, we are currently working on an Active Testing algorithm to do face detection and localization, searching for a face using any given number of feature spaces and any face classifier. The idea is to actively test the image for a face, computing the search strategy on the fly as information is gathered.

Currently working on this with Dr. Bruno Jedynak

Snake Curve Modeling

A fundamental understanding of the locomotion of living organisms is of great practical and scientific interest, and may lead to useful applications spanning fields such as robotics, medicine, and biology. One pertinent illustration is given with the nematode C. elegans. This small, 1 mm long, roundworm is widely used as a model system for biological research; its genome has been completely sequenced. For example, phenotypes of C. elegans motility are frequently used to identify genes involved in muscle function and model aspects of human Muscular Dystrophy. However, traditional methods for motility assessment have been driven largely by qualitative assays based on observation. For this reason, we are developing a series of image processing algorithms to facilitate quantitative motility analysis.

Currently working on this with Dr. Josue Sznitman at the Dept. of Mechanical and Aerospace Engineering at Princeton University.

Improving Cat Detection

This work had for objective to build a classifier for images of cats, and focused on particularly difficult images. Two approaches were implemented; one was using the Chamfer distance metric for matching, and the other learned models of our training images and used the Kullback-Leibler divergence as a metric of image comparison. A report on the implementation specifications and the results of the experimentations are presented in this report.

Worked on this at the IDIAP Research Institute with Dr. Francois Fleuret

Indian Driving Computer Vision Challenge

Given a Youtube video of a busy street intersection, our task was to detect, track, segment, recognize and count as many moving objects as possible. The task is a difficult one, as the camera is moving, the resolution is low, many objects are moving at once and creating occlusions. This was a competition project for a computer vision class. Our approach won 1st prize and results can be viewed here, with a report on the methods used here

Project was done with Giancarlo Troni

Teaching

I have been the TA for the following classes: 600.461 Computer Vision
600.363 Introduction to Algorithms
600.226 Data Structures
Some presentations I have done: Here are some slides I made for a two hour lecture on Viola and Jones Face Detection and Adaboost. Here is an additional derivation of Adaboost.
Here is a short presentation on object detection and recognition for a Freshman introduction class.