• 2002
• Domain
• Project link

To what extent can information visualization scale? This project designs new algorithms and interaction techniques to adapt popular information visualization representations, such as treemaps and scatter plot diagrams, for displaying one million of items or more in a effective way, without resorting to any aggregation technique.

To achieve this goal, we are developing special techniques to: experiment with non standard visual attributes such as shading, transparency and stereo-vision; use animation to help understanding view changes; experiment with new interaction techniques for dynamic labeling (extensions of excentric labels) and; animated "tours" to quickly explore a data set with several different views.

To manage optimally the screen real-estate, we are relying on hardware graphics acceleration to allow for smooth transitions between views, interpolation between layouts and synthesis of graphics attributes such as "overlaps" (among other things).

Description of Unique Features: * Visualizing One Million Items * Visualizing one million of items on a 1600x1200 screen is a challenge in term of visualization, graphics, perception and interaction. We have designed new techniques to achieve it for treemaps and scatter plots.

Visual Perception or Design Principles Applied: Visualization systems typically draw items with a one-pixel border, spending two lines and two columns of pixels and sending the geometry twice. We use slightly shaded quadrilaterals so that they remain distinguishable when tiled or stacked.

Data Analysis Techniques Applied: Dynamic Queries Dynamic queries rely on interactively filtering and redisplaying a data set through a continuous interaction. Current systems use "range-sliders" to filter one attribute either changing the smallest value, the largest, or sweeping a range of values between the smallest and the largest. To achieve the redisplay speed required for smooth interaction, we have designed a technique that relies on hardware acceleration.

The data set should be loaded into main memory. When the user activates a slider to perform the query dynamically, all the items are sent to the GPU and stored in a display list. The Z coordinate is calculated according to the attribute being filtered by the dynamic query so, for example, if a film database is displayed and the user wants to filter on the size of the film, the size is assigned to the Z-axis.

Each time the slider moves, a new near or far plane value is computed and sent to the GPU and the list is redisplayed, leaving the visibility computation to the hardware. Sliders or range-sliders are used to control the interaction. On current systems, their precision is limited to their size, augmenting the size increasing the precision at the cost of screen real-estate and longer movements to reach the slider. Our sliders and range-sliders are small (around 100 pixels) but their precision increase when the mouse leave their region. Fast changes can be done by keeping the mouse on the control's region while precision is achieved by going farther away from the control. See the animation (32Mb file).

Spatial Layout Techniques Applied: Combining software and hardware techniques provides a sustained performance around 2.5 million quads per second. By using texture mapping for animating treemaps, we achieve 10 frames per second for animating across any family of treemap. For scatter plots, we have only reached 3 frames per second for animations on 1 million items and 6 frames per second as worst for dynamic queries. Finding techniques for improving that speed would be useful but the next generation of graphics cards and computers will solve the problem.