#### DMCA

## Tuner: Principled parameter finding for image segmentation algorithms using visual response surface exploration (2011)

Venue: | IEEE Transactions on Visualization and Computer Graphics |

Citations: | 22 - 6 self |

### Citations

4675 | A Computational Approach to Edge Detection
- Canny
- 1986
(Show Context)
Citation Context ... parameters. For instance it is quite common to have thresholding parameters. One of the fundamental image processing algorithms is edge detection; the most popular algorithm, the Canny edge detector =-=[10]-=-, uses three parameters, controlling the size of a Gaussian smoothing function and thresholding with hysteresis (using min/max thresholds). In addition, almost any segmentation algorithm also includes... |

3948 | Snakes: Active contour models
- Kass, Witkin, et al.
- 1988
(Show Context)
Citation Context ...eighting of the importance of every energy term. The final segmentation φ̂ is obtained by minimizing (Eq. 1) as follows: φ̂ = argmin φ E(φ , I). (2) For example, consider the popular Snakes algorithm =-=[21]-=-. Here an approximate boundary evolves to the desired boundary guided by minimizing two competing energy terms. A boundary energy term attracts the solution to pixels with high gradients. However, sin... |

2120 | R.: Fast approximate energy minimization via graph cuts
- Boykov, Veksler, et al.
(Show Context)
Citation Context ...2 Algorithmic Parameters Algorithmic parameters fine-tune different parts of the algorithms. For variational segmentation, for instance, there exist approaches, like graph based approaches graph cuts =-=[7]-=- and random walker [16], where the energy term itself contains parameters to be tuned. We could describe these terms using E(φ , I,σ) where σ would impact how similar two nodes are that are connected ... |

906 |
A comparison of three methods for selecting values of input variables in the analysis of output from a computer code
- McKay, Beckman, et al.
- 1979
(Show Context)
Citation Context ...time the user can generate additional sample points in order to build up a more accurate model (i). As an alternative we provide the user two alternative sampling strategies: Latin Hypercube sampling =-=[28]-=- and random sampling. Both of these strategies place an exact number of samples in the parameter space. This allows the user to accurately interpret the running time. We show the generated sample poin... |

718 |
Measures of the amount of ecologic associations between spcies."
- Dice
- 1945
(Show Context)
Citation Context ...ed, but fine subtleties or 3D images are hard to inspect properly. Therefore, a number of other quality metrics have been developed. One of the most popular metrics is the Dice similarity coefficient =-=[14]-=-. It measures the overlap between a segmented region and ground truth, with a value of one corresponding to a perfect overlap. Precision and Recall are two other widely used quantities to assess the q... |

498 |
The Design and Analysis of Computer Experiments,
- 8Santner, Williams, et al.
- 2003
(Show Context)
Citation Context ...lished area of research in statistics employing this idea is known as DACE - the Design and Analysis of Computer Experiments. For a good introduction we refer the reader to the book by Santner et al. =-=[36]-=-. The particular model we are using has been well described by Jones et al. [20]. It has been successfully employed in a variety of computer experiments such as an ocean circulation model [15], a haza... |

481 |
Efficient global optimization of expensive black-box functions
- Jones, Schonlau, et al.
- 1998
(Show Context)
Citation Context ...e the stability of these regions. The contributions of this paper can be summarized as follows: (i) We develop a systematic model to explore the full parameter space based on a Gaussian process model =-=[20]-=-. (ii) We allow the user to visually explore the full parameter space using sliced-based navigation (similar to HyperSlice [43]) of the response function for up-to tens of parameters. (iii) We allow t... |

443 | SANKUR B.: Survey over image thresholding techniques and quantitative performance evaluation.
- SEZGIN
- 2004
(Show Context)
Citation Context ...respect to different criteria. Other commonly used tradeoffs include likelihood versus prior in Bayesian methods [1], loss versus penalty in pattern recognition [47] and image thresholding algorithms =-=[39]-=-. Our proposed tool is able to work with any segmentation technique controlled by a set of parameters and associated with a set of numerical quality measures. 2.4 Finding the Right Parameters Given a ... |

438 |
Elements of graphing data.
- Cleveland
- 1985
(Show Context)
Citation Context ...rameter space. 3 RELATED WORK Understanding and analyzing high-dimensional spaces has always been a challenge in statistical graphics as well as visualization. Approaches such as scatterplot matrices =-=[12]-=-, parallel coordinates [18], and star-glyphs [44] are now common for visualizing highdimensional data. Their main purpose, is to understand “point-clouds” or discrete entities. However, if one needs t... |

387 | Random walks for image segmentation.
- Grady
- 2006
(Show Context)
Citation Context ...rs Algorithmic parameters fine-tune different parts of the algorithms. For variational segmentation, for instance, there exist approaches, like graph based approaches graph cuts [7] and random walker =-=[16]-=-, where the energy term itself contains parameters to be tuned. We could describe these terms using E(φ , I,σ) where σ would impact how similar two nodes are that are connected through an edge. Algori... |

323 |
The plane with parallel coordinates
- Inselberg
- 1985
(Show Context)
Citation Context ...RK Understanding and analyzing high-dimensional spaces has always been a challenge in statistical graphics as well as visualization. Approaches such as scatterplot matrices [12], parallel coordinates =-=[18]-=-, and star-glyphs [44] are now common for visualizing highdimensional data. Their main purpose, is to understand “point-clouds” or discrete entities. However, if one needs to understand continuous hig... |

302 |
On the experimental attainment of optimum conditions
- Box, Wilson
- 1951
(Show Context)
Citation Context ... neighborhood around a single high-dimensional parameter combination. This has been improved with the authors current work [5]. 3.1 Statistics Our technique was born out of the work by Box and Wilson =-=[6]-=- in 1951. Their method is to fit gradually more and more complex estimating models to a complex function. One well-established area of research in statistics employing this idea is known as DACE - the... |

262 |
Current methods in medical image segmentation
- Pham, Xu, et al.
- 2001
(Show Context)
Citation Context ...gions with specific physiological activity in the case of functional imaging. In recent years a variety of semi- and fully automatic techniques have been developed to address the segmentation problem =-=[32]-=-. However, even the current state-of-the-art approaches fall short of providing a “silver bullet” for image segmentation. This has several reasons. One reason is that given some image, the segmentatio... |

247 | Design Galleries: A General Approach to Setting Parameters for Computer Graphics and Animation
- Marks, Andalman, et al.
- 1997
(Show Context)
Citation Context ... the availability of any ground truth. Therefore, their system is quite different from ours. Alternative approaches to facilitate parameter explorations have resulted in systems like Design Galleries =-=[25]-=-, Image Graphs [23], spreadsheet-like exploration interfaces [19], and VisTrails [37]. None of these approaches is utilizing an optimization function nor is the user able to see a comprehensive overvi... |

234 |
XmdvTool: Integrating multiple methods for visualzing multivariate data.
- Ward
- 1994
(Show Context)
Citation Context ...nalyzing high-dimensional spaces has always been a challenge in statistical graphics as well as visualization. Approaches such as scatterplot matrices [12], parallel coordinates [18], and star-glyphs =-=[44]-=- are now common for visualizing highdimensional data. Their main purpose, is to understand “point-clouds” or discrete entities. However, if one needs to understand continuous high-dimensional spaces, ... |

229 | Discriminative random fields: a discriminative framework for contextual interaction in classification.
- Kumar, Hebert
- 2003
(Show Context)
Citation Context ...ter Tuning in Image Segmentation Few papers tackled the automatic parameter estimation in image segmentation as the main focus are effective segmentation models and efficient algorithms. Kumar et al. =-=[22]-=- applied a pseudo-likelihood technique to estimate the parameters of a conditional random field algorithm. Szummer et al. [40] applied graph cuts to do maximum margin efficient learning of the segment... |

140 | ColorBrewer.org: An Online Tool for Selecting Colour Schemes for Maps.
- Harrower, Brewer
- 2003
(Show Context)
Citation Context ...color maps for the different response values in order to visually distinguish them. Different response variables are distinguished through different hues using quantitative colormaps from Colorbrewer =-=[17]-=-. We provide three different plot types of the data to the user. An example image of each is shown in Fig 4. Each of these plot types support a different task that the users want to perform. Response ... |

119 | Probabilistic Sensitivity Analysis of Complex Models: A Bayesian Approach,”
- Oakely, O’Hagan
- 2004
(Show Context)
Citation Context ...tatisticians is to fit the Gaussian process model and then evaluate the results of a variance decomposition. Two approaches to this method are discussed by Schonlau et al. [38] and Oakley and O’Hagan =-=[31]-=-. Our approach is to provide insight by viewing and interactively exploring the full response space. 3.2 Automatic Parameter Tuning in Image Segmentation Few papers tackled the automatic parameter est... |

104 | Learning CRFs using graph cuts. In:
- Szummer, Kohli, et al.
- 2008
(Show Context)
Citation Context ... are effective segmentation models and efficient algorithms. Kumar et al. [22] applied a pseudo-likelihood technique to estimate the parameters of a conditional random field algorithm. Szummer et al. =-=[40]-=- applied graph cuts to do maximum margin efficient learning of the segmentation parameters. Mcintosh and Hamarneh [26] optimized a non-convex energy function to find the optimal parameters. They later... |

58 | Querying and creating visualizations by analogy.
- Scheidegger, Vo, et al.
- 2007
(Show Context)
Citation Context ...m ours. Alternative approaches to facilitate parameter explorations have resulted in systems like Design Galleries [25], Image Graphs [23], spreadsheet-like exploration interfaces [19], and VisTrails =-=[37]-=-. None of these approaches is utilizing an optimization function nor is the user able to see a comprehensive overview of which places of the parameter space have been “looked at” and which have not. T... |

54 | Hyperslice: visualization of scalar functions of many variables.
- Wijk, Liere
- 1993
(Show Context)
Citation Context ...el to explore the full parameter space based on a Gaussian process model [20]. (ii) We allow the user to visually explore the full parameter space using sliced-based navigation (similar to HyperSlice =-=[43]-=-) of the response function for up-to tens of parameters. (iii) We allow the user to study the trade-off of up to two quality measures. (iv) We provide uncertainty visualization of the response surface... |

50 | Image graphs - a novel approach to visual data exploration.
- Ma
- 1999
(Show Context)
Citation Context ...f any ground truth. Therefore, their system is quite different from ours. Alternative approaches to facilitate parameter explorations have resulted in systems like Design Galleries [25], Image Graphs =-=[23]-=-, spreadsheet-like exploration interfaces [19], and VisTrails [37]. None of these approaches is utilizing an optimization function nor is the user able to see a comprehensive overview of which places ... |

48 | Where do people draw lines?
- Cole, Golovinskiy, et al.
- 2008
(Show Context)
Citation Context ...microtubule centerlines manually (the ground truth) for the test data and compare the automatically computed centerlines to these. The measurement we use is similar to the one utilized by Cole et al. =-=[13]-=- and results in numerical measures Precision and Recall which measure the estimated error rate in the algorithmic segmentation. Note that using Precision and Recall as measures requires that the groun... |

37 | Uncertainty-aware exploration of continuous parameter spaces using multivariate prediction
- Berger, Piringer, et al.
- 2011
(Show Context)
Citation Context ...ng an optimization function nor is the user able to see a comprehensive overview of which places of the parameter space have been “looked at” and which have not. The inspiring work by Piringer et al. =-=[33, 4, 5]-=- is perhaps closest to our work. Their system, HyperMoVal, was one of the first comprehensive environments for studying the impact of parameters on simulation experiments. HyperMoVal was also using th... |

34 | Result-driven exploration of simulation parameter spaces for visual effects design
- Bruckner, Möller
(Show Context)
Citation Context ...he parameters at the start of a new simulation with the intention of optimizing the final output according to some quality measure. This is closer to the work by Bruckner and Möller in FluidExplorer =-=[9]-=-. However, one of the major accomplishments of FluidExplorer was dealing with simulation outputs where temporal behavior is crucial. Furthermore, they did not address optimization of any objective fun... |

33 | Continuous scatterplots
- Bachthaler, Weiskopf
- 2008
(Show Context)
Citation Context ...f one needs to understand continuous high-dimensional spaces, these approaches fail mostly, since they do not properly convey the continuity of the underlying space. Recently, Bachthaler and Weiskopf =-=[2]-=- extended scatterplots in order to properly portray continuous functions. However, the essence of scatterplots and similar approaches is to separate the data values from its intrinsic embedding in som... |

24 |
Adaptive Markov Modeling for Mutual-Information-Based, Unsupervised MRI Brain-Tissue Classification”. Medical Image Analysis,
- Tasdizen
- 2006
(Show Context)
Citation Context ...these different measures simultaneously to ensure better segmentation performance with respect to different criteria. Other commonly used tradeoffs include likelihood versus prior in Bayesian methods =-=[1]-=-, loss versus penalty in pattern recognition [47] and image thresholding algorithms [39]. Our proposed tool is able to work with any segmentation technique controlled by a set of parameters and associ... |

22 | HyperMoVal: Interactive visual validation of regression models for real-time simulation
- Piringer, Berger, et al.
- 2010
(Show Context)
Citation Context ...ng an optimization function nor is the user able to see a comprehensive overview of which places of the parameter space have been “looked at” and which have not. The inspiring work by Piringer et al. =-=[33, 4, 5]-=- is perhaps closest to our work. Their system, HyperMoVal, was one of the first comprehensive environments for studying the impact of parameters on simulation experiments. HyperMoVal was also using th... |

21 |
A spreadsheet interface for visualization exploration. Visualization ’00
- Jankun-Kelly, Ma
(Show Context)
Citation Context ... quite different from ours. Alternative approaches to facilitate parameter explorations have resulted in systems like Design Galleries [25], Image Graphs [23], spreadsheet-like exploration interfaces =-=[19]-=-, and VisTrails [37]. None of these approaches is utilizing an optimization function nor is the user able to see a comprehensive overview of which places of the parameter space have been “looked at” a... |

19 | Is a Single Energy Functional Sufficient? Adaptive Energy Functionals and Automatic Initialization
- McIntosh, Hamarneh
- 2007
(Show Context)
Citation Context ...o estimate the parameters of a conditional random field algorithm. Szummer et al. [40] applied graph cuts to do maximum margin efficient learning of the segmentation parameters. Mcintosh and Hamarneh =-=[26]-=- optimized a non-convex energy function to find the optimal parameters. They later extended their technique using a constrained convex energy function to avoid sensitivity to the initial parameter set... |

19 | Visualization of parameter space for image analysis
- Pretorius, Bray, et al.
- 2011
(Show Context)
Citation Context ...e user to express their preference after comparing different simulation outputs (see e.g. the work by Brochu et al. [8]). This is an area also known as active learning. Very recently Pretorius et al. =-=[34]-=- have been developing a system for the exploration of parameter values for image segmentation. In their case they are not making use of any quality measures and don’t assume the availability of any gr... |

17 | Stagewise Lasso
- Zhao, Yu
- 2007
(Show Context)
Citation Context ...e better segmentation performance with respect to different criteria. Other commonly used tradeoffs include likelihood versus prior in Bayesian methods [1], loss versus penalty in pattern recognition =-=[47]-=- and image thresholding algorithms [39]. Our proposed tool is able to work with any segmentation technique controlled by a set of parameters and associated with a set of numerical quality measures. 2.... |

17 | Automatic particle selection: results of a comparative study,
- Zhu, Carragher, et al.
- 2004
(Show Context)
Citation Context ... Precision and Recall as measures requires that the ground truth segmentation is complete, since missing lines would result in a low Precision even for a perfect automatic segmentation. Previous work =-=[48]-=-, albeit from a different domain, leads us to believe that setting the Precision in our case to above 0.9 and Recall to above 0.8 will yield trustworthy results. Tuning parameters according to these c... |

15 | A bayesian interactive optimization approach to procedural animation design.
- Brochu, Brochu, et al.
- 2010
(Show Context)
Citation Context ...ive function. Approaches where an objective function is missing typically require the user to express their preference after comparing different simulation outputs (see e.g. the work by Brochu et al. =-=[8]-=-). This is an area also known as active learning. Very recently Pretorius et al. [34] have been developing a system for the exploration of parameter values for image segmentation. In their case they a... |

14 |
Kinetic modeling based probabilistic segmentation for molecular images,”
- Saad, Hamarneh, et al.
- 2008
(Show Context)
Citation Context ...the process of applying mathematical models to analyze the temporal tracer activity in order to extract clinically or experimentally relevant information. An extension of the algorithm by Saad et al. =-=[35]-=- is being developed by adding more energy terms to make the segmentation more robust. The goal is to segment 2D+Time PET images into six functional regions: background, skull, grey matter, white matte... |

13 |
Parameter space exploration of an ocean general circulation model using an isopycnal mixing parameterization.J.
- Gough, Welch
- 1994
(Show Context)
Citation Context ... et al. [36]. The particular model we are using has been well described by Jones et al. [20]. It has been successfully employed in a variety of computer experiments such as an ocean circulation model =-=[15]-=-, a hazard-effect model for volcano eruption prediction [3], and an arctic sea ice simulation [11]. However, the typical approach by statisticians is to fit the Gaussian process model and then evaluat... |

13 | World lines
- Waser, Fuchs, et al.
(Show Context)
Citation Context ...alization community. Computational steering has been a known problem for a while, which is addressed by several researchers in the visualization community, most recently in WorldLines by Waser et al. =-=[45]-=-. This problem is fundamentally different from the problem we are trying to solve. In computational steering the user studies a simulation over time and actually wants to change the parameters while t... |

12 |
The prosection matrix, a tool to support the interactive exploration of statistical models and data
- Tweedie, Spence
- 1998
(Show Context)
Citation Context ...ional function had been suggested by van Wijk and van Liere using a technique which they coined Hyperslice [43]. This idea was extended by Tweedie and Spence by what they called the Prosection Matrix =-=[42]-=-. Here, a thick (n−2)-dimensional slab is being summarized as opposed to a simple 2D slice of the n-dimensional space under study. Since we believe, that a slice will be a more accurate portrayal of t... |

11 | Arctic sea ice variability: Model sensitivities and a multidecadal simulation
- Chapman, Welch, et al.
- 1994
(Show Context)
Citation Context ...as been successfully employed in a variety of computer experiments such as an ocean circulation model [15], a hazard-effect model for volcano eruption prediction [3], and an arctic sea ice simulation =-=[11]-=-. However, the typical approach by statisticians is to fit the Gaussian process model and then evaluate the results of a variance decomposition. Two approaches to this method are discussed by Schonlau... |

10 |
An Extension of Wilkinson’s Algorithm for Positioning Tick Labels on Axes
- Talbot, Lin, et al.
(Show Context)
Citation Context ...he R environment, mlegp and lhs respectively. The link between Scala and R is provided by the JRI Java library. The axes labels were determined using the algorithm and code described in Talbot et al. =-=[41]-=-. 8 LESSONS LEARNED AND FUTURE WORK Because of enthusiastic feedback by our collaborators and their colleagues, we plan to extend the capabilities of our tool and make it available to a wider audience... |

7 | Using Statistical and Computer Models to Quantify Volcanic Hazards
- Bayarri, Berger, et al.
- 2009
(Show Context)
Citation Context ...l described by Jones et al. [20]. It has been successfully employed in a variety of computer experiments such as an ocean circulation model [15], a hazard-effect model for volcano eruption prediction =-=[3]-=-, and an arctic sea ice simulation [11]. However, the typical approach by statisticians is to fit the Gaussian process model and then evaluate the results of a variance decomposition. Two approaches t... |

7 | Optimal weights for convex functionals in medical image segmentation
- McIntosh, Hamarneh
(Show Context)
Citation Context ...ized a non-convex energy function to find the optimal parameters. They later extended their technique using a constrained convex energy function to avoid sensitivity to the initial parameter settings =-=[27]-=-. The common goal in these algorithms is to learn the parameters so that the ground truth emerges as the optimal solution. This is achieved by optimizing yet another energy function. Therefore, these ... |

7 | Screening the Input Variables to a Computer Model Via Analysis of Variance and Visualization
- Schonlau, Welch
- 2006
(Show Context)
Citation Context ...r, the typical approach by statisticians is to fit the Gaussian process model and then evaluate the results of a variance decomposition. Two approaches to this method are discussed by Schonlau et al. =-=[38]-=- and Oakley and O’Hagan [31]. Our approach is to provide insight by viewing and interactively exploring the full response space. 3.2 Automatic Parameter Tuning in Image Segmentation Few papers tackled... |

4 |
Interactive visual analysis of multiobjective optimizations
- Berger, Piringer
- 2010
(Show Context)
Citation Context ...ng an optimization function nor is the user able to see a comprehensive overview of which places of the parameter space have been “looked at” and which have not. The inspiring work by Piringer et al. =-=[33, 4, 5]-=- is perhaps closest to our work. Their system, HyperMoVal, was one of the first comprehensive environments for studying the impact of parameters on simulation experiments. HyperMoVal was also using th... |

4 |
Automated tracing of microtubules in electron tomograms of plastic embedded samples of Caenorhabditis elegans embryos
- Weber, Greenan, et al.
- 2012
(Show Context)
Citation Context ...d cell division. The segmentation algorithm proceeds by first enhancing the centerlines of the microtubules in the tomograms by computing normalized cross-correlation with an idealized cylinder [24], =-=[46]-=-. In the second step the geometry of the centerlines is reconstructed using a simple iterative greedy traversal in the enhanced volume. The tracing algorithm has three parameters: a threshold, minCorr... |

3 |
NAUTILUS method: An interactive technique in multiobjective optimization based on the nadir point
- Miettinen, Eskelinen, et al.
- 2010
(Show Context)
Citation Context ... An example use case is examining Dice coefficients for different structures simultaneously. There are approaches in the literature, that deal with the trade-off of multiple objective functions [42], =-=[29]-=-. We believe their work can help in improving Tuner. Further, there is a need to compare the tuning of parameters for multiple images from different patients or multiple images from different noise le... |

1 |
editor. Cellular Electron Microscopy
- MacIntosh
- 2007
(Show Context)
Citation Context ...ing and cell division. The segmentation algorithm proceeds by first enhancing the centerlines of the microtubules in the tomograms by computing normalized cross-correlation with an idealized cylinder =-=[24]-=-, [46]. In the second step the geometry of the centerlines is reconstructed using a simple iterative greedy traversal in the enhanced volume. The tracing algorithm has three parameters: a threshold, m... |