Vlad I. Morariu

I am a postdoctoral research associate at the University of Maryland, associated with the Computer Vision Lab and the Institute for Advanced Computer Studies. I received my doctoral degree in Computer Science at the University of Maryland, College Park, with Prof. Larry Davis as my advisor. I completed my B.S. and M.S. degrees at Penn State University (with Prof. Octavia Camps as my research advisor). My research interests are in Computer Vision, Pattern Recognition, Machine Learning, and Artificial Intelligence.

Research Projects

Analyzing Activities Involving Interacting People
I am currently working on analyzing activities, such as basketball, that involve multiple interacting people. This work entails low and mid level analysis, such as tracking humans and recognizing actions, as well as high level reasoning that incorporates external knowledge about the world.

Tracking People's Hands and Feet Using Mixed Network AND/OR Search

We describe a framework that leverages mixed probabilistic and deterministic networks and their AND/OR search space to efficiently find and track the hands and feet of multiple interacting humans in 2D from a single camera view. Our framework detects and tracks multiple people's heads, hands, and feet through partial or full occlusion; requires few constraints (does not require multiple views, high image resolution, knowledge of performed activities, or large training sets); and makes use of constraints and AND/OR Branch-and-Bound with lazy evaluation and carefully computed bounds to efficiently solve the complex network that results from the consideration of inter-person occlusion. Our main contributions are 1) a multi-person part-based formulation that emphasizes extremities and allows for the globally optimal solution to be obtained in each frame, and 2) an efficient and exact optimization scheme that relies on AND/OR Branch-and-Bound, lazy factor evaluation, and factor cost sensitive bottom-up bound computation.

Automatic Tuning for Fast Gaussian Summation
We provide an algorithm that combines tree methods with the Improved Fast Gauss Transform (IFGT). As originally proposed the IFGT suffers from two problems: (1) the Taylor series expansion does not perform well for very low bandwidths, and (2) parameter selection is not trivial and can drastically affect performance and ease of use. We address the first problem by employing a tree data structure, and the second problem by using an online tuning approach that results in a black box method that automatically chooses the evaluation method and its parameters to yield the best performance for the input data, desired accuracy, and bandwidth. In addition, the new IFGT parameter selection approach allows for tighter error bounds. Our approach chooses the fastest method at negligible additional cost, and has superior performance in comparisons with previous approaches.

Representing Visibility Context for Action Understanding

We proposed a representation of visibility/spatial context based on visibility features (obtained from isovists and visibility graphs) that is suitable for human action understanding. Using a Bayes net, we then used our visibility context representation to reason about 2-dimensional trajectories (top view) generated by an agent performing a simple search-based task in various layouts. Human subjects were asked to interpret the trajectories 1) to demonstrate that knowledge of visibility context improves interpretation of our task and 2) to provide a baseline against which our algorithm can be compared. Our framework was able to match the performance of humans.

Appearance Modeling for Multi-camera Correspondence and Tracking
(M.S. research)
We learned generative appearance models by extracting object appearance from single or multiple views and learning its evolution on a manifold over time. Using target dynamics, we were able to predict future appearance in each view. In the multiple view case we learned correspondences between the implicit low-dimensional representation of each high-dimensional object view by either aligning low-dimensional coordinates during nonlinear manifold learning, or learning the dynamics of how low-dimensional coordinates in separate views evolved together over time. Our model allowed us to "hallucinate" the appearance of occluded targets by 1) predicting future appearance in each view temporally or by 2) predicting the appearance in one view given the appearance in another.

Publications

Sameh Khamis, Vlad I. Morariu, and Larry S. Davis. A Flow Model for Joint Action Recognition and Identity Maintenance. IEEE Conference on Computer on Computer Vision and Pattern Recognition (CVPR), 2012. PDF BibTeX

Ryan Farrell, Om Oza, Ning Zhang, Vlad I. Morariu, Trevor Darrell, and Larry S. Davis. Subordinate Categorization Using Volumetric Primitives and Pose-Normalized Appearance. IEEE International Conference on Computer Vision (ICCV), 2011. [ORAL] PDF BibTeX

Vlad I. Morariu and Larry S. Davis. Multi-agent event recognition in structured scenarios. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2011. PDF BibTeX Appendix Video

Vlad I. Morariu, David Harwood, and Larry S. Davis. Tracking People's Hands and Feet Using Mixed Network AND/OR Search. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), submitted, 2010.

Vlad I. Morariu, Balaji Vasan Srinivasan, Vikas C. Raykar, Ramani Duraiswami, and Larry S. Davis. Automatic online tuning for fast Gaussian summation. Advances in Neural Information Processing Systems (NIPS), 2008. PDF BibTeX

Vlad I. Morariu, V. Shiv Naga Prasad, and Larry S. Davis. Human Activity Understanding using Visibility Context. IEEE/RSJ IROS Workshop: From sensors to human spatial concepts (FS2HSC), 2007. PDF BibTeX

Benjamin Fransen, Vlad Morariu, Eric Martinson, Samuel Blisard, Matthew Marge, Scott Thomas, Alan Schultz, and Dennis Perzanowski. Using Vision, Acoustics, and Natural Language for Disambiguation. IEEE International Conference on Human-Robot Interaction (HRI) 2007. PDF BibTeX

Vlad I. Morariu, Octavia I. Camps, Mario Sznaier, and Hwasup Lim. Robust Cooperative Visual Tracking: A Combined Nonlinear Dimensionality Reduction/Robust Identification Approach. In Advances in Cooperative Control and Optimization, M. Hirsch, R. Murphey, P. Pardalos and D. Grundel, Eds., Springer Verlag, 2007. PDF BibTeX

Vlad I. Morariu and Octavia I. Camps. Modeling Correspondences in Multi-camera Tracking using Nonlinear Manifold Learning and Target Dynamics. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2006. PDF BibTeX

Hwasup Lim, Vlad I. Morariu, Octavia I. Camps, and Mario Sznaier. Dynamic Appearance Modeling for Human Tracking. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2006. PDF BibTeX

Software

FIGTree - Fast Improved Gauss Transform with Tree Data Structure (FIGTree Homepage)
A library for fast computation of Gauss transforms in multiple dimensions, using the Improved Fast Gauss Transform and Approximate Nearest Neighbor searching. The nearest neighbor searching is performed using the ANN library, available at http://www.cs.umd.edu/~mount/ANN/. This software allows for efficient computation of probabilities by Kernel Density Estimation (KDE), and can reduce complexity of algorithms commonly used in Computer Vision, Machine Learning, etc, that must evaluate the Gauss transform. The publication describing the newest improvements in the code is the NIPS 2008 paper by Morariu et al. Previous publications related to this approach are provided on Vikas Raykar's page.
NOTE: A new version of the code based on the NIPS 2008 paper has been released! Now FIGTree can be used as a black box.