The Visualization Toolkit (VTK) is a widely used software system for data processing and visualization. It is used in scientific computing, medical image analysis, computational geometry, rendering, image processing and informatics. In this chapter we provide a brief overview of VTK, including some of the basic design patterns that make it a successful system.

To really understand a software system it is essential to not only understand what problem it solves, but also the particular culture in which it emerged. In the case of VTK, the software was ostensibly developed as a 3D visualization system for scientific data. But the cultural context in which it emerged adds a significant back story to the endeavor, and helps explains why the software was designed and deployed as it was.

At the time VTK was conceived and written, its initial authors (Will Schroeder, Ken Martin, Bill Lorensen) were researchers at GE Corporate R&D. We were heavily invested in a precursor system known as LYMB which was a Smalltalk-like environment implemented in the C programming language. While this was a great system for its time, as researchers we were consistently frustrated by two major barriers when trying to promote our work: 1) IP issues and 2) non-standard, proprietary software. IP issues were a problem because trying to distribute the software outside of GE was nearly impossible once the corporate lawyers became involved. Second, even if we were deploying the software inside of GE, many of our customers balked at learning a proprietary, non-standard system since the effort to master it did not transition with an employee once she left the company, and it did not have the widespread support of a standard tool set. Thus in the end the primary motivation for VTK was to develop an open standard, or collaboration platform through which we could easily transition technology to our customers. Thus choosing an open source license for VTK was probably the most important design decision that we made.

The final choice of a non-reciprocal, permissive license (i.e., BSD not GPL) in hindsight was an exemplary decision made by the authors because it ultimately enabled the service and consulting based business that became Kitware. At the time we made the decision we were mostly interested in reduced barriers to collaborating with academics, research labs, and commercial entities. We have since discovered that reciprocal licenses are avoided by many organizations because of the potential havoc they can wreak. In fact we would argue that reciprocal licenses do much to slow the acceptance of open source software, but that is an argument for another time. The point here is: one of the major design decisions to make relative to any software system is the choice of copyright license. It's important to review the goals of the project and then address IP issues appropriately.

24.1. What Is VTK?

VTK was initially conceived as a scientific data visualization system. Many people outside of the field naively consider visualization a particular type of geometric rendering: examining virtual objects and interacting with them. While this is indeed part of visualization, in general data visualization includes the whole process of transforming data into sensory input, typically images, but also includes tactile, auditory, and other forms. The data forms not only consist of geometric and topological constructs, including such abstractions as meshes or complex spatial decompositions, but attributes to the core structure such as scalars (e.g., temperature or pressure), vectors (e.g., velocity), tensors (e.g., stress and strain) plus rendering attributes such as surface normals and texture coordinate.

Note that data representing spatial-temporal information is generally considered part of scientific visualization. However there are more abstract data forms such as marketing demographics, web pages, documents and other information that can only be represented through abstract (i.e., non-spatial temporal) relationships such as unstructured documents, tables, graphs, and trees. These abstract data are typically addressed by methods from information visualization. With the help of the community, VTK is now capable of both scientific and information visualization.

As a visualization system, the role of VTK is to take data in these forms and ultimately transform them into forms comprehensible by the human sensory apparatus. Thus one of the core requirements of VTK is its ability to create data flow pipelines that are capable of ingesting, processing, representing and ultimately rendering data. Hence the toolkit is necessarily architected as a flexible system and its design reflects this on many levels. For example, we purposely designed VTK as a toolkit with many interchangeable components that can be combined to process a wide variety of data.

24.2. Architectural Features

Before getting too far into the specific architectural features of VTK, there are high-level concepts that have significant impact on developing and using the system. One of these is VTK's hybrid wrapper facility. This facility automatically generates language bindings to Python, Java, and Tcl from VTK's C++ implementation (additional languages could be and have been added). Most high-powered developers will work in C++. User and application developers may use C++ but often the interpreted languages mentioned above are preferred. This hybrid compiled/interpreted environment combines the best of both worlds: high performance compute-intensive algorithms and flexibility when prototyping or developing applications. In fact this approach to multi-language computing has found favor with many in the scientific computing community and they often use VTK as a template for developing their own software.

In terms of software process, VTK has adopted CMake to control the build; CDash/CTest for testing; and CPack for cross-platform deployment. Indeed VTK can be compiled on almost any computer including supercomputers which are often notoriously primitive development environments. In addition, web pages, wiki, mailing lists (user and developer), documentation generation facilities (i.e., Doxygen) and a bug tracker (Mantis) round out the development tools.

24.2.1. Core Features

As VTK is an object-oriented system, the access of class and instance data members is carefully controlled in VTK. In general, all data members are either protected or private. Access to them is through Set and Get methods, with special variations for Boolean data, modal data, strings and vectors. Many of these methods are actually created by inserting macros into the class header files. So for example:


become on expansion:

virtual void SetTolerance(double);
virtual double GetTolerance();

There are many reasons for using these macros beyond simply code clarity. In VTK there are important data members controlling debugging, updating an object's modified time (MTime), and properly managing reference counting. These macros correctly manipulate these data and their use is highly recommended. For example, a particularly pernicious bug in VTK occurs when the object's MTime is not managed properly. In this case code may not execute when it should, or may execute too often.

One of the strengths of VTK is its relatively simplistic means of representing and managing data. Typically various data arrays of particular types (e.g., vtkFloatArray) are used to represent contiguous pieces of information. For example, a list of three XYZ points would be represented with a vtkFloatArray of nine entries (x,y,z, x,y,z, etc.) There is the notion of a tuple in these arrays, so a 3D point is a 3-tuple, whereas a symmetric 3×3 tensor matrix is represented by a 6-tuple (where symmetry space savings are possible). This design was adopted purposely because in scientific computing it is common to interface with systems manipulating arrays (e.g., Fortran) and it is much more efficient to allocate and deallocate memory in large contiguous chunks. Further, communication, serializing and performing IO is generally much more efficient with contiguous data. These core data arrays (of various types) represent much of the data in VTK and have a variety of convenience methods for inserting and accessing information, including methods for fast access, and methods that automatically allocate memory as needed when adding more data. Data arrays are subclasses of the vtkDataArray abstract class meaning that generic virtual methods can be used to simplify coding. However, for higher performance static, templated functions are used which switch based on type, with subsequent, direct access into the contiguous data arrays.

In general C++ templates are not visible in the public class API; although templates are used widely for performance reasons. This goes for STL as well: we typically employ the PIMPL1 design pattern to hide the complexities of a template implementation from the user or application developer. This has served us particularly well when it comes to wrapping the code into interpreted code as described previously. Avoiding the complexity of the templates in the public API means that the VTK implementation, from the application developer point of view, is mostly free of the complexities of data type selection. Of course under the hood the code execution is driven by the data type which is typically determined at run time when the data is accessed.

Some users wonder why VTK uses reference counting for memory management versus a more user-friendly approach such as garbage collection. The basic answer is that VTK needs complete control over when data is deleted, because the data sizes can be huge. For example, a volume of byte data 1000×1000×1000 in size is a gigabyte in size. It is not a good idea to leave such data lying around while the garbage collector decides whether or not it is time to release it. In VTK most classes (subclasses of vtkObject) have the built-in capability for reference counting. Every object contains a reference count that it initialized to one when the object is instantiated. Every time a use of the object is registered, the reference count is increased by one. Similarly, when a use of the object is unregistered (or equivalently the object is deleted) the reference count is reduced by one. Eventually the object's reference count is reduced to zero, at which point it self destructs. A typical example looks like the following:

vtkCamera *camera = vtkCamera::New();   //reference count is 1
camera->Register(this);                 //reference count is 2
camera->Unregister(this);               //reference count is 1
renderer->SetActiveCamera(camera);      //reference count is 2
renderer->Delete();                     //ref count is 1 when renderer is deleted
camera->Delete();                       //camera self destructs

There is another important reason why reference counting is important to VTK—it provides the ability to efficiently copy data. For example, imagine a data object D1 that consists of a number of data arrays: points, polygons, colors, scalars and texture coordinates. Now imagine processing this data to generate a new data object D2 which is the same as the first plus the addition of vector data (located on the points). One wasteful approach is to completely (deep) copy D1 to create D2, and then add the new vector data array to D2. Alternatively, we create an empty D2 and then pass the arrays from D1 to D2 (shallow copy), using reference counting to keep track of data ownership, finally adding the new vector array to D2. The latter approach avoids copying data which, as we have argued previously, is essential to a good visualization system. As we will see later in this chapter, the data processing pipeline performs this type of operation routinely, i.e., copying data from the input of an algorithm to the output, hence reference counting is essential to VTK.

Of course there are some notorious problems with reference counting. Occasionally reference cycles can exist, with objects in the cycle referring to each other in a mutually supportive configuration. In this case, intelligent intervention is required, or in the case of VTK, the special facility implemented in vtkGarbageCollector is used to manage objects which are involved in cycles. When such a class is identified (this is anticipated during development), the class registers itself with the garbage collector and overloads its own Register and UnRegister methods. Then a subsequent object deletion (or unregister) method performs a topological analysis on the local reference counting network, searching for detached islands of mutually referencing objects. These are then deleted by the garbage collector.

Most instantiation in VTK is performed through an object factory implemented as a static class member. The typical syntax appears as follows:

vtkLight *a = vtkLight::New();

What is important to recognize here is what is actually instantiated may not be a vtkLight, it could be a subclass of vtkLight (e.g., vtkOpenGLLight). There are a variety of motivations for the object factory, the most important being application portability and device independence. For example, in the above we are creating a light in a rendered scene. In a particular application on a particular platform, vtkLight::New may result in an OpenGL light, however on different platforms there is potential for other rendering libraries or methods for creating a light in the graphics system. Exactly what derived class to instantiate is a function of run-time system information. In the early days of VTK there were a myriad of options including gl, PHIGS, Starbase, XGL, and OpenGL. While most of these have now vanished, new approaches have appeared including DirectX and GPU-based approaches. Over time, an application written with VTK has not had to change as developers have derived new device specific subclasses to vtkLight and other rendering classes to support evolving technology. Another important use of the object factory is to enable the run-time replacement of performance-enhanced variations. For example, a vtkImageFFT may be replaced with a class that accesses special-purpose hardware or a numerics library.

24.2.2. Representing Data

One of the strengths of VTK is its ability to represent complex forms of data. These data forms range from simple tables to complex structures such as finite element meshes. All of these data forms are subclasses of vtkDataObject as shown in Figure 24.1 (note this is a partial inheritance diagram of the many data object classes).

[Data Object Classes]

Figure 24.1: Data Object Classes

One of the most important characteristics of vtkDataObject is that it can be processed in a visualization pipeline (next subsection). Of the many classes shown, there are just a handful that are typically used in most real world applications. vtkDataSet and derived classes are used for scientific visualization (Figure 24.2). For example, vtkPolyData is used to represent polygonal meshes; vtkUnstructuredGrid to represent meshes, and vtkImageData represents 2D and 3D pixel and voxel data.

[Data Set Classes]

Figure 24.2: Data Set Classes

24.2.3. Pipeline Architecture

VTK consists of several major subsystems. Probably the subsystem most associated with visualization packages is the data flow/pipeline architecture. In concept, the pipeline architecture consists of three basic classes of objects: objects to represent data (the vtkDataObjects discussed above), objects to process, transform, filter or map data objects from one form into another (vtkAlgorithm); and objects to execute a pipeline (vtkExecutive) which controls a connected graph of interleaved data and process objects (i.e., the pipeline). Figure 24.3 depicts a typical pipeline.

[Typical Pipeline]

Figure 24.3: Typical Pipeline

While conceptually simple, actually implementing the pipeline architecture is challenging. One reason is that the representation of data can be complex. For example, some datasets consist of hierarchies or grouping of data, so executing across the data requires non-trivial iteration or recursion. To compound matters, parallel processing (whether using shared-memory or scalable, distributed approaches) require partitioning data into pieces, where pieces may be required to overlap in order to consistently compute boundary information such as derivatives.

The algorithm objects also introduce their own special complexity. Some algorithms may take multiple inputs and/or produce multiple outputs of different types. Some can operate locally on data (e.g., compute the center of a cell) while others require global information, for example to compute a histogram. In all cases, the algorithms treat their inputs as immutable, algorithms only read their input in order to produce their output. This is because data may be available as input to multiple algorithms, and it is not a good idea for one algorithm to trample on the input of another.

Finally the executive can be complicated depending on the particulars of the execution strategy. In some cases we may wish to cache intermediate results between filters. This minimizes the amount of recomputation that must be performed if something in the pipeline changes. On the other hand, visualization data sets can be huge, in which case we may wish to release data when it is no longer needed for computation. Finally, there are complex execution strategies, such as multi-resolution processing of data, which require the pipeline to operate in iterative fashion.

To demonstrate some of these concepts and further explain the pipeline architecture, consider the following C++ example:

vtkPExodusIIReader *reader = vtkPExodusIIReader::New();

vtkContourFilter *cont = vtkContourFilter::New();
cont->SetValue(0, 200);

vtkQuadricDecimation *deci = vtkQuadricDecimation::New();
deci->SetTargetReduction( 0.75 );

vtkXMLPolyDataWriter *writer = vtkXMLPolyDataWriter::New();

In this example, a reader object reads a large unstructured grid (or mesh) data file. The next filter generates an isosurface from the mesh. The vtkQuadricDecimation filter reduces the size of the isosurface, which is a polygonal dataset, by decimating it (i.e., reducing the number of triangles representing the isocontour). Finally after decimation the new, reduced data file is written back to disk. The actual pipeline execution occurs when the Write method is invoked by the writer (i.e., upon demand for the data).

As this example demonstrates, VTK's pipeline execution mechanism is demand driven. When a sink such as a writer or a mapper (a data rendering object) needs data, it asks its input. If the input filter already has the appropriate data, it simply returns the execution control to the sink. However, if the input does not have the appropriate data, it needs to compute it. Consequently, it must first ask its input for data. This process will continue upstream along the pipeline until a filter or source that has "appropriate data" or the beginning of the pipeline is reached, at which point the filters will execute in correct order and the data will flow to the point in the pipeline at which it was requested.

Here we should expand on what "appropriate data" means. By default, after a VTK source or filter executes, its output is cached by the pipeline in order to avoid unnecessary executions in the future. This is done to minimize computation and/or I/O at the cost of memory, and is configurable behavior. The pipeline caches not only the data objects but also the metadata about the conditions under which these data objects were generated. This metadata includes a time stamp (i.e., ComputeTime) that captures when the data object was computed. So in the simplest case, the "appropriate data" is one that was computed after all of the pipeline objects upstream from it were modified. It is easier to demonstrate this behavior by considering the following examples. Let's add the following to the end of the previous VTK program:

vtkXMLPolyDataWriter *writer2 = vtkXMLPolyDataWriter::New();

As explained previously, the first writer->Write call causes the execution of the entire pipeline. When writer2->Write() is called, the pipeline will realize that the cached output of the decimation filter is up to date when it compares the time stamp of the cache with the modification time of the decimation filter, the contour filter and the reader. Therefore, the data request does not have to propagate past writer2. Now, let's consider the following change.

cont->SetValue(0, 400);

vtkXMLPolyDataWriter *writer2 = vtkXMLPolyDataWriter::New();

Now the pipeline executive will realize that the contour filter was modified after the outputs of the contour and decimation filters were last executed. Thus, the cache for these two filters are stale and they have to be re-executed. However, since the reader was not modified prior to the contour filter its cache is valid and hence the reader does not have to re-execute.

The scenario described here is the simplest example of a demand-driven pipeline. VTK's pipeline is much more sophisticated. When a filter or a sink requires data, it can provide additional information to request specific data subsets. For example, a filter can perform out-of-core analysis by streaming pieces of data. Let's change our previous example to demonstrate.

vtkXMLPolyDataWriter *writer = vtkXMLPolyDataWriter::New();



Here the writer asks the upstream pipeline to load and process data in two pieces each of which are streamed independently. You may have noticed that the simple execution logic described previously will not work here. By this logic when the Write function is called for the second time, the pipeline should not re-execute because nothing upstream changed. Thus to address this more complex case, the executives have additional logic to handle piece requests such as this. VTK's pipeline execution actually consists of multiple passes. The computation of the data objects is actually the last pass. The pass before then is a request pass. This is where sinks and filters can tell upstream what they want from the forthcoming computation. In the example above, the writer will notify its input that it wants piece 0 of 2. This request will actually propagate all the way to the reader. When the pipeline executes, the reader will then know that it needs to read a subset of the data. Furthermore, information about which piece the cached data corresponds to is stored in the metadata for the object. The next time a filter asks for data from its input, this metadata will be compared with the current request. Thus in this example the pipeline will re-execute in order to process a different piece request.

There are several more types of request that a filter can make. These include requests for a particular time step, a particular structured extent or the number of ghost layers (i.e., boundary layers for computing neighborhood information). Furthermore, during the request pass, each filter is allowed to modify requests from downstream. For example, a filter that is not able to stream (e.g., the streamline filter) can ignore the piece request and ask for the whole data.

24.2.4. Rendering Subsystem

At first glance VTK has a simple object-oriented rendering model with classes corresponding to the components that make up a 3D scene. For example, vtkActors are objects that are rendered by a vtkRenderer in conjunction with a vtkCamera, with possibly multiple vtkRenderers existing in a vtkRenderWindow. The scene is illuminated by one or more vtkLights. The position of each vtkActor is controlled by a vtkTransform, and the appearance of an actor is specified through a vtkProperty. Finally, the geometric representation of an actor is defined by a vtkMapper. Mappers play an important role in VTK, they serve to terminate the data processing pipeline, as well as interface to the rendering system. Consider this example where we decimate data and write the result to a file, and then visualize and interact with the result by using a mapper:

vtkOBJReader *reader = vtkOBJReader::New();

vtkTriangleFilter *tri = vtkTriangleFilter::New();

vtkQuadricDecimation *deci = vtkQuadricDecimation::New();
deci->SetTargetReduction( 0.75 );

vtkPolyDataMapper *mapper = vtkPolyDataMapper::New();

vtkActor *actor = vtkActor::New();

vtkRenderer *renderer = vtkRenderer::New();

vtkRenderWindow *renWin = vtkRenderWindow::New();

vtkRenderWindowInteractor *interactor = vtkRenderWindowInteractor::New();


Here a single actor, renderer and render window are created with the addition of a mapper that connects the pipeline to the rendering system. Also note the addition of a vtkRenderWindowInteractor, instances of which capture mouse and keyboard events and translate them into camera manipulations or other actions. This translation process is defined via a vtkInteractorStyle (more on this below). By default many instances and data values are set behind the scenes. For example, an identity transform is constructed, as well as a single default (head) light and property.

Over time this object model has become more sophisticated. Much of the complexity has come from developing derived classes that specialize on an aspect of the rendering process. vtkActors are now specializations of vtkProp (like a prop found on stage), and there are a whole slew of these props for rendering 2D overlay graphics and text, specialized 3D objects, and even for supporting advanced rendering techniques such as volume rendering or GPU implementations (see Figure 24.4).

Similarly, as the data model supported by VTK has grown, so have the various mappers that interface the data to the rendering system. Another area of significant extension is the transformation hierarchy. What was originally a simple linear 4×4 transformation matrix, has become a powerful hierarchy that supports non-linear transformations including thin-plate spline transformation. For example, the original vtkPolyDataMapper had device-specific subclasses (e.g., vtkOpenGLPolyDataMapper). In recent years it has been replaced with a sophisticated graphics pipeline referred to as the "painter" pipeline illustrated in Figure 24.4.

[Display Classes]

Figure 24.4: Display Classes

The painter design supports a variety of techniques for rendering data that can be combined to provide special rendering effects. This capability greatly surpasses the simple vtkPolyDataMapper that was initially implemented in 1994.

Another important aspect of a visualization system is the selection subsystem. In VTK there is a hierarchy of "pickers", roughly categorized into objects that select vtkProps based on hardware-based methods versus software methods (e.g., ray-casting); as well as objects that provide different levels of information after a pick operations. For example, some pickers provide only a location in XYZ world space without indicating which vtkProp they have selected; others provide not only the selected vtkProp but a particular point or cell that make up the mesh defining the prop geometry.

24.2.5. Events and Interaction

Interacting with data is an essential part of visualization. In VTK this occurs in a variety of ways. At its simplest level, users can observe events and respond appropriately through commands (the command/observer design pattern). All subclasses of vtkObject maintain a list of observers which register themselves with the object. During registration, the observers indicate which particular event(s) they are interested in, with the addition of an associated command that is invoked if and when the event occurs. To see how this works, consider the following example in which a filter (here a polygon decimation filter) has an observer which watches for the three events StartEvent, ProgressEvent, and EndEvent. These events are invoked when the filter begins to execute, periodically during execution, and then on completion of execution. In the following the vtkCommand class has an Execute method that prints out the appropriate information relative to the time it take to execute the algorithm:

class vtkProgressCommand : public vtkCommand
    static vtkProgressCommand *New() { return new vtkProgressCommand; }
    virtual void Execute(vtkObject *caller, unsigned long, void *callData)
      double progress = *(static_cast<double*>(callData));
      std::cout << "Progress at " << progress<< std::endl;

vtkCommand* pobserver = vtkProgressCommand::New();

vtkDecimatePro *deci = vtkDecimatePro::New();
deci->SetInputConnection( byu->GetOutputPort() );
deci->SetTargetReduction( 0.75 );
deci->AddObserver( vtkCommand::ProgressEvent, pobserver );

While this is a primitive form of interaction, it is a foundational element to many applications that use VTK. For example, the simple code above can be easily converted to display and manage a GUI progress bar. This Command/Observer subsystem is also central to the 3D widgets in VTK, which are sophisticated interaction objects for querying, manipulating and editing data and are described below.

Referring to the example above, it is important to note that events in VTK are predefined, but there is a back door for user-defined events. The class vtkCommand defines the set of enumerated events (e.g., vtkCommand::ProgressEvent in the above example) as well as a user event. The UserEvent, which is simply an integral value, is typically used as a starting offset value into a set of application user-defined events. So for example vtkCommand::UserEvent+100 may refer to a specific event outside the set of VTK defined events.

From the user's perspective, a VTK widget appears as an actor in a scene except that the user can interact with it by manipulating handles or other geometric features (the handle manipulation and geometric feature manipulation is based on the picking functionality described earlier.) The interaction with this widget is fairly intuitive: a user grabs the spherical handles and moves them, or grabs the line and moves it. Behind the scenes, however, events are emitted (e.g., InteractionEvent) and a properly programmed application can observe these events, and then take the appropriate action. For example they often trigger on the vtkCommand::InteractionEvent as follows:

vtkLW2Callback *myCallback = vtkLW2Callback::New();
  myCallback->PolyData = seeds;    // streamlines seed points, updated on interaction
  myCallback->Actor = streamline;  // streamline actor, made visible on interaction

vtkLineWidget2 *lineWidget = vtkLineWidget2::New();

VTK widgets are actually constructed using two objects: a subclass of vtkInteractorObserver and a subclass of vtkProp. The vtkInteractorObserver simply observes user interaction in the render window (i.e., mouse and keyboard events) and processes them. The subclasses of vtkProp (i.e., actors) are simply manipulated by the vtkInteractorObserver. Typically such manipulation consists of modifying the vtkProp's geometry including highlighting handles, changing cursor appearance, and/or transforming data. Of course, the particulars of the widgets require that subclasses are written to control the nuances of widget behavior, and there are more than 50 different widgets currently in the system.

24.2.6. Summary of Libraries

VTK is a large software toolkit. Currently the system consists of approximately 1.5 million lines of code (including comments but not including automatically generated wrapper software), and approximately 1000 C++ classes. To manage the complexity of the system and reduce build and link times the system has been partitioned into dozens of subdirectories. Table 24.1 lists these subdirectories, with a brief summary describing what capabilities the library provides.

Common core VTK classes
Filtering classes used to manage pipeline dataflow
Rendering rendering, picking, image viewing, and interaction
VolumeRendering volume rendering techniques
Graphics 3D geometry processing
GenericFiltering non-linear 3D geometry processing
Imaging imaging pipeline
Hybrid classes requiring both graphics and imaging functionality
Widgets sophisticated interaction
IO VTK input and output
Infovis information visualization
Parallel parallel processing (controllers and communicators)
Wrapping support for Tcl, Python, and Java wrapping
Examples extensive, well-documented examples

Table 24.1: VTK Subdirectories

24.3. Looking Back/Looking Forward

VTK has been an enormously successful system. While the first line of code was written in 1993, at the time of this writing VTK is still growing strong and if anything the pace of development is increasing.2 In this section we talk about some lessons learned and future challenges.

24.3.1. Managing Growth

One of the most surprising aspects to the VTK adventure has been the project's longevity. The pace of development is due to several major reasons:

  • New algorithms and capabilities continue to be added. For example, the informatics subsystem (Titan, primarily developed by Sandia National Labs and Kitware) is a recent significant addition. Additional charting and rendering classes are also being added, as well as capabilities for new scientific dataset types. Another important addition were the 3D interaction widgets. Finally, the on-going evolution of GPU-based rendering and data processing is driving new capabilities in VTK.
  • The growing exposure and use of VTK is a self-perpetuating process that adds even more users and developers to the community. For example, ParaView is the most popular scientific visualization application built on VTK and is highly regarded in the high-performance computing community. 3D Slicer is a major biomedical computing platform that is largely built on VTK and received millions of dollars per year in funding.
  • VTK's development process continues to evolve. In recent years the software process tools CMake, CDash, CTest, and CPack have been integrated into the VTK build environment. More recently, the VTK code repository has moved to Git and a more sophisticated work flow. These improvements ensure that VTK remains on the leading edge of software development in the scientific computing community.

While growth is exciting, validates the creation of the software system, and bodes well for the future of VTK, it can be extremely difficult to manage well. As a result, the near term future of VTK focuses more on managing the growth of the community as well as the software. Several steps have been taken in this regard.

First, formalized management structures are being created. An Architecture Review Board has been created to guide the development of the community and technology, focusing on high-level, strategic issues. The VTK community is also establishing a recognized team of Topic Leads to guide the technical development of particular VTK subsystems.

Next, there are plans to modularize the toolkit further, partially in response to workflow capabilities introduced by git, but also to recognize that users and developers typically want to work with small subsystems of the toolkit, and do not want to build and link against the entire package. Further, to support the growing community, it's important that contributions of new functionality and subsystems are supported, even if they are not necessarily part of the core of the toolkit. By creating a loose, modularized collection of modules it is possible to accommodate the large number of contributions on the periphery while maintaining core stability.

24.3.2. Technology Additions

Besides the software process, there are many technological innovations in the development pipeline.

  • Co-processing is a capability where the visualization engine is integrated into the simulation code, and periodically generates data extracts for visualization. This technology greatly reduces the need to output large amounts of complete solution data.
  • The data processing pipeline in VTK is still too complex. Methods are under way to simplify and refactor this subsystem.
  • The ability to directly interact with data is increasingly popular with users. While VTK has a large suite of widgets, many more interaction techniques are emerging including touch-screen-based and 3D methods. Interaction will continue its development at a rapid pace.
  • Computational chemistry is increasing in importance to materials designers and engineers. The ability to visualize and interact with chemistry data is being added to VTK.
  • The rendering system in VTK has been criticized for being too complex, making it difficult to derive new classes or support new rendering technology. In addition, VTK does not directly support the notion of a scene graph, again something that many users have requested.
  • Finally new forms of data are constantly emerging. For example, in the medical field hierarchical volumetric datasets of varying resolution (e.g., confocal microscopy with local magnification).

24.3.3. Open Science

Finally Kitware and more generally the VTK community are committed to Open Science. Pragmatically this is a way of saying we will promulgate open data, open publication, and open source—the features necessary to ensure that we are creating reproducible scientific systems. While VTK has long been distributed as an open source and open data system, the documentation process has been lacking. While there are decent books [Kit10,SML06] there have been a variety of ad hoc ways to collect technical publications including new source code contributions. We are improving the situation by developing new publishing mechanisms like the VTK Journal3 that enable of articles consisting of documentation, source code, data, and valid test images. The journal also enables automated reviews of the code (using VTK's quality software testing process) as well as human reviews of the submission.

24.3.4. Lessons Learned

While VTK has been successful there are many things we didn't do right:

  • Design Modularity: We did a good job choosing the modularity of our classes. For example, we didn't do something as silly as creating an object per pixel, rather we created the higher-level vtkImageClass that under the hood treats data arrays of pixel data. However in some cases we made our classes too high level and too complex, in many instances we've had to refactor them into smaller pieces, and are continuing this process. One prime example is the data processing pipeline. Initially, the pipeline was implemented implicitly through interaction of the data and algorithm objects. We eventually realized that we had to create an explicit pipeline executive object to coordinate the interaction between data and algorithms, and to implement different data processing strategies.
  • Missed Key Concepts: Once of our biggest regrets is not making widespread use of C++ iterators. In many cases the traversal of data in VTK is akin to the scientific programming language Fortran. The additional flexibility of iterators would have been a significant benefit to the system. For example, it is very advantageous to process a local region of data, or only data satisfying some iteration criterion.
  • Design Issues: Of course there is a long list of design decisions that are not optimal. We have struggled with the data execution pipeline, having gone through multiple generations each time making the design better. The rendering system too is complex and hard to derive from. Another challenge resulted from the initial conception of VTK: we saw it as a read-only visualization system for viewing data. However, current customers often want it to be capable of editing data, which requires significantly different data structures.

One of the great things about an open source system like VTK is that many of these mistakes can and will be rectified over time. We have an active, capable development community that is improving the system every day and we expect this to continue into the foreseeable future.


  2. See the latest VTK code analysis at