a system for more intuitive multivariate volume

2
A System for More Intuitive Multivariate Volume Exploration and Visualization Liang Zhou * SCI Institute and the School of Computing, University of Utah Charles Hansen SCI Institute and the School of Computing, University of Utah Figure 1: An overview of the proposed user interface. ABSTRACT We propose a novel linked-view multivariate volume exploration and visualization system that facilitates feature exploration and ex- traction for highly complex datasets. Through three interaction stages, namely: data probing, qualitative analysis and feature re- finement, the user is able to locate and extract features from com- plex multivariate datasets. Unlike previous linked-view systems, our proposed system offers a more intuitive and less laborious in- teraction for domain users. The system has been successfully uti- lized by geologists to extract features in complicated multivariate 3D seismic survey data. Keywords: Multivariate volume, transfer functions, user interface. 1 I NTRODUCTION Scientific data sets have been expanded not only in size but also dimensionality. As dimensionality increases due to multiple at- tributes, new interaction systems for exploratory visualization are needed. Use cases utilizing linked-view systems that allow the user to set value ranges or transfer functions (TFs) in a value domain editor and examine the classified result in the spatial domain [1, 4] are clearly shown with simulation datasets. However, extracting meaningful features in real world measurement datasets, e.g. multi- variate 3D seismic survey, via these systems is not trivial. 3D seis- mic imaging has been the standard for oil and gas exploration for decades, and more recently, multi-attribute volumes derived from the seismic amplitude volume have been used to aid the understand- ing of seismic surveys [2]. Features inside the seismic dataset have * e-mail: [email protected] e-mail: [email protected] to be recognized in the spatial domain by a geology expert, and the features have complicated combination of attribute values and subtle difference from their surroundings. Therefore, it is too labo- rious to extract features by iterating between TF design in the value domain and getting feedback from the results rendered in the spa- tial domain, especially when the dimensionality is high. To address these issues, we propose a novel multivariate volumetric visualiza- tion system that is more intuitive for the domain users and provides better classification results. 2 SYSTEM FRAMEWORK AND USER I NTERFACE The work flow of our proposed system is comprised of three major stages: data probing, qualitative analysis and feature refinement. Data probing is the process where the user discovers regions of in- terest by examining multivariate data slices. The regions of interest can be conveniently selected using lasso tool or ”magic wand” tool. Once the regions of interest are selected, a simple, yet efficient, voxel query operation that inquires the multivariate data values is performed. The user then performs a qualitative analysis, i.e. ex- tracting and rendering volumetric features by means of designing high-dimensional(HD) TFs or 2D TFs on dimensionally reduced space. Kernel density estimation [7] is utilized to automatically generate the TFs and to robustly discard outliers from the probed samples. The HD TFs can then be fine-tuned directly in the parallel coordinate plot based (PCP) TF editor. In addition, automatic 2D Gaussian TFs on the dimensionally reduced view offers an simpler alternative for more distinct features. On many occasions, however, different features share similar data values and thus an additional feature refinement stage is introduced to refine the features classi- fied by the TFs. Features are refined by the user via segmentation brushes or lassos which are applied directly on the volume render- ing view or the multi-panel view. The user interface of our system is seen in Figure 1 with four tightly linked views labeled. Multi-panel View: the multi-panel view simultaneously shows all attributes of a data slice with synchronized user interactions.

Upload: others

Post on 19-Dec-2021

6 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: A System for More Intuitive Multivariate Volume

A System for More Intuitive Multivariate Volume Exploration andVisualization

Liang Zhou∗

SCI Institute and the School of Computing,University of Utah

Charles Hansen†

SCI Institute and the School of Computing,University of Utah

Figure 1: An overview of the proposed user interface.

ABSTRACT

We propose a novel linked-view multivariate volume explorationand visualization system that facilitates feature exploration and ex-traction for highly complex datasets. Through three interactionstages, namely: data probing, qualitative analysis and feature re-finement, the user is able to locate and extract features from com-plex multivariate datasets. Unlike previous linked-view systems,our proposed system offers a more intuitive and less laborious in-teraction for domain users. The system has been successfully uti-lized by geologists to extract features in complicated multivariate3D seismic survey data.

Keywords: Multivariate volume, transfer functions, user interface.

1 INTRODUCTION

Scientific data sets have been expanded not only in size but alsodimensionality. As dimensionality increases due to multiple at-tributes, new interaction systems for exploratory visualization areneeded. Use cases utilizing linked-view systems that allow the userto set value ranges or transfer functions (TFs) in a value domaineditor and examine the classified result in the spatial domain [1, 4]are clearly shown with simulation datasets. However, extractingmeaningful features in real world measurement datasets, e.g. multi-variate 3D seismic survey, via these systems is not trivial. 3D seis-mic imaging has been the standard for oil and gas exploration fordecades, and more recently, multi-attribute volumes derived fromthe seismic amplitude volume have been used to aid the understand-ing of seismic surveys [2]. Features inside the seismic dataset have

∗e-mail: [email protected]†e-mail: [email protected]

to be recognized in the spatial domain by a geology expert, andthe features have complicated combination of attribute values andsubtle difference from their surroundings. Therefore, it is too labo-rious to extract features by iterating between TF design in the valuedomain and getting feedback from the results rendered in the spa-tial domain, especially when the dimensionality is high. To addressthese issues, we propose a novel multivariate volumetric visualiza-tion system that is more intuitive for the domain users and providesbetter classification results.

2 SYSTEM FRAMEWORK AND USER INTERFACE

The work flow of our proposed system is comprised of three majorstages: data probing, qualitative analysis and feature refinement.Data probing is the process where the user discovers regions of in-terest by examining multivariate data slices. The regions of interestcan be conveniently selected using lasso tool or ”magic wand” tool.Once the regions of interest are selected, a simple, yet efficient,voxel query operation that inquires the multivariate data values isperformed. The user then performs a qualitative analysis, i.e. ex-tracting and rendering volumetric features by means of designinghigh-dimensional(HD) TFs or 2D TFs on dimensionally reducedspace. Kernel density estimation [7] is utilized to automaticallygenerate the TFs and to robustly discard outliers from the probedsamples. The HD TFs can then be fine-tuned directly in the parallelcoordinate plot based (PCP) TF editor. In addition, automatic 2DGaussian TFs on the dimensionally reduced view offers an simpleralternative for more distinct features. On many occasions, however,different features share similar data values and thus an additionalfeature refinement stage is introduced to refine the features classi-fied by the TFs. Features are refined by the user via segmentationbrushes or lassos which are applied directly on the volume render-ing view or the multi-panel view. The user interface of our systemis seen in Figure 1 with four tightly linked views labeled.

Multi-panel View: the multi-panel view simultaneously showsall attributes of a data slice with synchronized user interactions.

Page 2: A System for More Intuitive Multivariate Volume

To improve perception, each attribute can have a specifically de-signed color map whose dynamic range is changeable and can beco-rendered with a context attribute. The user can put lassos overregions of interest or brush inside a region of interest using ananisotropic diffusion segmentation based ’magic wand’ tool. Themultivariate data values of the samples collected inside the dataprober are efficiently queried via GPU-based conditional histogramcomputation. A joint conditional histogram jch(a,b) f of two at-tributes a and b is a 2D histogram showing the joint distribution ofattribute values Ya and Yb of voxels V whose evaluated result froma certain boolean function f (V, ⃗Y (V )) ( ⃗Y (V ) being the attribute val-ues of V ) is true. For a multivariate volume of N attributes, giventhe selected regions as a masking condition, a set of N−1 joint con-ditional histograms can be computed to record the query results.

High-Dimensional Transfer Function View: the HD TF viewdisplays the PCP and the pair-wise 2D histograms and allows directuser manipulation of the TF widgets in both plots. Specifically, thePCP and pair-wise 2D histograms of the entire data are renderedas context, while the plots of TF classified data or user queried dataare highlighted in front. Any updates to the TFs will invoke updatesof the plots via voxel query, and as such the user immediately getsfeedback in the data value space which is important to guide theexploration. Kernel density estimation is utilized to automaticallycreate HD TFs from the queried data samples represented by theconditional histograms. The outliers inside the samples can be dis-carded by thresholding the density field and a smooth opacity TFcan be created. To reduce the computational complexity, we sepa-rate the N dimensional value space into N −1 2D value spaces, i.e.a 2D+2D+ · · ·+2D (N −1 of 2D) space.

Projection View: Fastmap [3] is adopted to reduce the dimen-sionality of the original data down to 2D, and the histogram ofthe dimensionality reduced data is shown in the projection view.The projection view supports both traditional 2D TF widget edit-ing and automatic 2D Gaussian TF generation. Specifically, theuser selects samples in the multi-panel view to invoke the popu-lar Expectation Maximization optimization with Gaussian mixturemodel which automatically generates Gaussian TFs. This view fur-ther simplifies the TF design process and is especially useful fordistinct features and simulation datasets.

Volume Rendering View: a slice-based volume renderer thatutilizes a multivariate version of the directional occlusion shad-ing [6] is implemented. The directional occlusion shading over-comes the artifacts of traditional Phong shading with degeneratedgradients and aids the visualization of seismic datasets [5]. Oncethe volume classified by the TFs is rendered in the volume render-ing view, the user can refine the classified volume directly on thevolume renderer via either region growing segmentation or a selec-tion lasso.

3 USE CASE: 3D SEISMIC DATASET

The data used in this case is a part of the public 3D seismic sur-vey dataset ”New Zealand”. Five attributes are utilized to createthe visualization: the seismic amplitude(Amp), the median filteredcontinuous sculpture volume of horizons (Seg MedFilter), theinstantaneous amplitude (Inst Amp), the semblance (Semb), andthe thickness of the semblance (Semb Thick). Using the five at-tributes, our collaborating geologist is able to extract clear geolog-ical features as shown in Figure 2. The extracted geological fea-tures include: a shallow channel complex shown in red, a salt domeshown in yellow, a deeper channel shown in purple and a largestfault in green. These features are important in the oil industry sincepotential oil and gas reservoirs are likely to be inside them. Theexpert examines through slices in the multi-panel view to recognizefeatures of interest and then uses a combination of the lasso tool andthe ’magic wand’ tool to select them: lasso tool is used to select ex-act inhomogeneous regions while the ’magic wand’ tool is conve-

Figure 2: The rendering result of the extracted geological features.

nient for homogeneous region selection. It is easy to highlight thesalt dome with automatic 2D Gaussian TFs in the projection viewsince this feature has a distinct cluster in the dimensionally reducedspace. Other features, however, are more difficult to discern fromtheir surroundings in both the spatial and the value domain. Dueto the data loss of the dimensionality reduction procedure, we findthat the dimensionally reduced space cannot capture the subtle dif-ference of these features and their surroundings. Instead, the HDTF provides enough precision to extract these features. After theautomatic KDE based TFs are generated, it often requires the ex-pert to further fine tune the TFs in the HD TF view. For all thefeatures, the TFs alone cannot yield very clean and well separatedresults, therefore, a refinement stage has to be performed. Boththe region growing segmentation and the lasso selection are usedto refine these extracted features: the region growing segmentationbrush is used to extract isolated structure while the lasso selectionis used for picking structures with complicated neighborhood.

4 FUTURE WORK

In the future, we would like to improve our work two-fold: scal-ability and user interaction. Currently, the system supports onlymultivariate volumes that can fit into GPU memory. We are devel-oping a GPU-based out-of-core method for volume rendering, con-ditional histogram computation and parallel coordinate plot render-ing to support full size 3D seismic data. The user interaction of thesystem is intuitive, but is still time consuming for very complicateddata sets. By introducing advanced machine learning methods, wehope to make the interaction more automated and thus less labori-ous.

REFERENCES

[1] J. Blaas, C. Botha, and F. Post. Extensions of parallel coordinates forinteractive exploration of large multi-timepoint data sets. IEEE Trans-actions on Visualization and Computer Graphics, 14(6):1436 –1451,nov.-dec. 2008.

[2] S. Chopra and K. J. Marfurt. Seismic Attributes for Prospect ID andReservoir Characterization (Geophysical Developments No. 11). Soci-ety Of Exploration Geophysicists, 2007.

[3] C. Faloutsos and K.-I. Lin. Fastmap: a fast algorithm for indexing,data-mining and visualization of traditional and multimedia datasets.SIGMOD Rec., 24(2):163–174, May 1995.

[4] H. Guo, H. Xiao, and X. Yuan. Multi-dimensional transfer functiondesign based on flexible dimension projection embedded in parallel co-ordinates. In Pacific Visualization Symposium, 2011.

[5] D. Patel, S. Bruckner, I. Viola, and E. Groller. Seismic volume vi-sualization for horizon extraction. In Pacific Visualization Symposium(PacificVis), 2010 IEEE, pages 73 –80, march 2010.

[6] M. Schott, V. Pegoraro, C. D. Hansen, K. Boulanger, and K. Boua-touch. A directional occlusion shading model for interactive direct vol-ume rendering. Computer Graphics Forum, 28(3):855–862, 2009.

[7] B. W. Silverman. Density Estimation for Statistics and Data Analysis.Chapman & Hall/CRC, 1986.