Real Time Feature Extraction & Tracking

Overview for the distributed feature extraction and tracking system

    A full overview of both the feature extraction and feature tracking distributed processes are shown in the figure below. Each processor determines the features and the tracking information as the data becomes available. The data is either being computed by the simulation or it is being read in by the visualization program.  The viz accumulator is running where the actual visualization takes place, i.e. a local workstation.


Figure 1. Overview for the distributed feature extraction and tracking system

Distributed Feature Extraction

    In our previous work (pdf ps), distributed feature extraction algorithm and implementations are described. Figure 2 shows the "partial-merge" strategy for feature extraction.


     Figure 2. The "partial-merge" strategy for feature extraction.

Distributed Feature Tracking

 Two distributed feature extraction and tracking examples are shown in Figure 4 &5. The number below each image is the feature count of that block data. In the later time images, each object gets the same color as its matched feature of its previous timestep.


Figure 4. Parallel feature extraction and tracking example.


Figure 5. Another parallel feature extraction and tracking example

GrACE Implementation & DISCOVER Portal

GrACE
    In order for the feature extraction and tracking routines to run in-situ, the implementation must utilize the same data structure as the ongoing simulation so as to not incur any data copying overhead. For this implementation, we used the GrACE infrastructure.

   The implementation contains three parts:

A in-situ feature extraction & tracking example is shown in Figure 6. The upper part shows the feature tracking results for 5 timesteps of the RM3D** simulation. Each feature is given its own color and child features inherit the color from the parent. The lower graph shows how the volume of the red feature changes over time. Note how the feature splits. In this example, the DISCOVER portal was used to control the simulation and start the feature tracking at timestep 1000.


Figure 6. In-situ feature extraction and tracking.

DISCOVER

    DISCOVER is a generic framework that enables interactive steering of scientific applications and also allows for collaborative visualization of data sets generated by simulations. DISCOVER is supported by a suite of detachable interfaces and analysis modules and allows users to interact with, interrogate, control and steer GrACE-based applications through a web based portal. A snapshot of the GrACE interface portal is shown in Figure 7.


Figure 7. In-situ feature extraction and tracking with DISCOVER platform


** The RM3D simualtion uses a compressible Euler equation for shock accelerated inhomogeneous flows (Richtmyer-Meshkov) . In this environment, baroclinic vorticity is playing the major role of hydrodynamic instability and late time turbulent mixing. 8000 timesteps were run with a resolution of 256x64x64, with a threshold of 50%.
 
This work ( PDF   PS)will be presented at High-Performance Computing Symposium 2003, Orlando, FL, March 31th - April 2nd, 2003(part of the SCS Advanced Simulation Technologies Conference, March 30 - April 3rd, 2003).