recent brews

  • Multiple Instance Filtering for Object Tracking

    Tracking an object whose appearance is constantly evolving requires that a relevant model be maintained. We view this from the perspective of a joint regression and classification problem and devise a multi-part object representation that enables us to join the ideas of robust filtering and multiple instance learning.

  • Diving

    Can we learn to quantify action quality? A challenging new task for action research. An initial classification analysis was presented at the ACCV2010 VECTaR workshop.

  • Analysis of Human Actions

    We view human actions from the perspective of time series. We show that by focusing on the temporal characteristics of actions very low dimensional features can be highly discriminative. Our work presents several techniques for building dictionaries of action primitives from time series and explores simple classification algorithms that leverage this representation. A second line of work explores dynamic models of motion, decomposing actions into a dynamic system that remains constant over the course of observation (properties of actor) and a set of sparse inputs (action signature).

  • Towards Keyword Based Category Recognition

    Can we automatically gather good training images from the web for any object keyword with no prior knowledge? One might imagine existing image search engines as a natural starting point. Unfortunately current image search techniques do not utilize image information, thus returned results typically contain many irrelevant images. We explore a non-parametric measure of strangeness in the space of holistic image descriptors to remove visually inconsistent search results.

a taste of the past

  • Simultaneous Localization and Mapping (SLAM) with Dynamic Object Tracking

    I did some evaluation work of SLAM algorithms and developed a Matlab version of FastSLAM with several variations. Tracking of dynamic objects was incorporated for robustness in dynamic environments. I also put together a simple self-contained robot simulator for visualization and testing.

  • Highly Organized Bunch Of Scavengers (HOBOS):
    A Multi Agent Approach to Vision Based Robot Scavenging

    Together with my Vision Lab colleagues, Brian Fulkerson and Jeremi Sudol, we designed and implemented a system for multi agent robot scavenging. The architecture of the system is specifically engineered for scalability in the number of agents and robustness to long periods of network partitioning. Each agent can operate entirely on its own, but cooperates with others even when sparse communication is available.

    I worked primarily on the localization and communication elements of the system, which utilized structure from motion (SFM), hidden markov models (HMMs), scale invariant feature transform (SIFT), and delay tolerant networking concepts from peer-to-peer mobile ad-hoc networking.

  • Dense 3D Mapping with Monocular Vision

    While at Harvey Mudd I constructed a large C++ toolkit for building 3D maps from sequential images captured by a moving robot. The toolkit also included graphics tools to display the intermediate results of each computational step in the pipeline. This was the project that first got me into vision and led me to where I currently stand.