### Citations

12867 | The nature of statistical learning theory
- Vapnik
- 1995
(Show Context)
Citation Context ...y flat, indicating that performance is insensitive to the number of bins. 4.3. Support Vector Machine We presented the feature vectors to SVM light [13], a C implementation of Support Vector Machines =-=[14]-=-. The SVM attempts to find a multidimensional cutting plane that separates the positive (trail) and negative (nontrail) examples. We experimented with the four available kernel options: linear, polyno... |

11683 | Maximum likelihood from incomplete data via the EM algorithm
- Dempster, Laird, et al.
- 1977
(Show Context)
Citation Context ...f trail given the node, P(v | n) is a Gaussian distribution for that node, and P(n) is the node prior. Model parameters are learned from the training data using the expectation maximization algorithm =-=[17]-=-. We verified that the performance of the model is relatively robust to a wide range for the number of nodes, N. For the results in this paper we used 200 nodes. 5. Trail Extraction In the final step ... |

10431 | Introduction to Algorithms
- Cormen, Leiserson, et al.
- 2001
(Show Context)
Citation Context ...te the global minimum for the objective function: N ∑ f ( path ) = P ( not _ trail ) (2) 1 If we let N vary we can find the above minimum efficiently with a shortest path algorithm such as Dijkstra’s =-=[18]-=-, and doing so leads to a baseline algorithm which we test below. However, this favors longer paths since we are, in effect, maximizing P(trail) along the path. Figure 3 is an example of an image wher... |

3873 | Snakes: Active contour models
- Kass, Witkin, et al.
- 1987
(Show Context)
Citation Context ...s can be used to pre-process the GPS data, placing it on trails in the neighborhood of the data. The GPS-snakes method draws inspiration from the active contour (snakes) model developed by Kass et al =-=[8]-=-, but it exploits properties of GPS data and trail image appearance to correct GPS tracks. In Figure 2, original GPS data is shown along with the GPSsnakes corrected output. Table 1 shows the positive... |

3721 | Normalized cuts and image segmentation - Shi, Malik |

3533 |
Equations of state calculations by fast computing machines
- Metropolis, Rosenbluth, et al.
- 1953
(Show Context)
Citation Context ... to explore the trail space, we allow acceptance of worse proposals according to the ratio: t f ( x ) < f ( x') α (5) Where α is drawn from U(0,1). Though similar to the Metropolis Hastings algorithm =-=[19]-=- [20], our objective function (4) is not a probability distribution, so we are not using Metropolis Hastings as such. The ratio (5) is reversed since we seek the minimum of our objective function, rat... |

2033 |
Monte Carlo sampling methods using Markov chains and their applications
- Hastings
- 1970
(Show Context)
Citation Context ...xplore the trail space, we allow acceptance of worse proposals according to the ratio: t f ( x ) < f ( x') α (5) Where α is drawn from U(0,1). Though similar to the Metropolis Hastings algorithm [19] =-=[20]-=-, our objective function (4) is not a probability distribution, so we are not using Metropolis Hastings as such. The ratio (5) is reversed since we seek the minimum of our objective function, rather t... |

652 | Matching words and pictures
- Barnard, Duygulu, et al.
- 2003
(Show Context)
Citation Context ...M light software provides soft classification which we normalized to use as a crude estimate of probability. 4.4. Multi-modal mixture model We also trained a generative multi-modal mixture model [15] =-=[16]-=- to classify image points. Here we assume that an image pixel and its label (“trail” or “not_trail”) are concurrently generated as follows. First, a hidden factor, or node, is chosen according to a pr... |

553 |
Learning to classify text using support vector machines
- Joachims
- 2002
(Show Context)
Citation Context ...Tucson Mountain Park dataset. The plot was relatively flat, indicating that performance is insensitive to the number of bins. 4.3. Support Vector Machine We presented the feature vectors to SVM light =-=[13]-=-, a C implementation of Support Vector Machines [14]. The SVM attempts to find a multidimensional cutting plane that separates the positive (trail) and negative (nontrail) examples. We experimented wi... |

429 | An Analysis of Bayesian Classifiers
- Langley, Iba, et al.
- 1992
(Show Context)
Citation Context ...’ assumption that all of the feature vector components are independent of each other. This is not the case with our vectors, but even when this assumption is violated, naïve Bayes often performs well =-=[12]-=-. We generate two sets of histograms, one for trail and one for not trail. Among each set there is a histogram for each component of the feature vector. With the histograms computed, the likelihood of... |

390 | Preattentive texture discrimination with early vision mechanisms - Malik, Perona - 1990 |

186 | An Active Testing Model for Tracking Roads in Satellite Images
- Geman, Jedynak
- 1996
(Show Context)
Citation Context ...hborhoods. To determine if pixels are “road” or “background” operators such as edge detectors, ridge finders, crest detectors, morphological operators and even road-specific operators have been tried =-=[4]-=-. Figure 1: Example USGS aerial photograph, from the Santa Rita mountains Road tracking [4] extends the local information in plausible directions. However, to keep the search space reasonable, the ass... |

125 |
Detection of roads and linear structures in low resolution aerial imagery using a multisource knowledge integration technique
- Fischler, Tenenbaum, et al.
- 1981
(Show Context)
Citation Context ...at roads generally change direction slowly is made—a very reasonable simplification in the road domain, but clearly not prudent in the case of trails. Our approach is closer to that by Fischler et al =-=[5]-=-, which uses a dynamic programming optimization over the pixel lattice. The cost function is based on weights determined by local operators. To extend this work to trail data would require non-trivial... |

26 | Survey of Work on Road Extraction in Aerial and Satellite Images
- Auclair-Fortierm, Ziou, et al.
- 2000
(Show Context)
Citation Context ...e of previous work focused on finding recreational trails in remote sensed images. However, the problem of finding roads from such data has been well studied, with a variety of approaches surveyed in =-=[3]-=-. We briefly review some of what has been learned from work on roads below. However, we emphasize that trails are more difficult than roads since they are narrower, less predictable, less uniform, mor... |

18 | A computational model of texture segmentation - Malik, Perona - 1989 |

13 |
Learning the Semantics of Words and
- Barnard, Forsyth
- 2001
(Show Context)
Citation Context ...he SVM light software provides soft classification which we normalized to use as a crude estimate of probability. 4.4. Multi-modal mixture model We also trained a generative multi-modal mixture model =-=[15]-=- [16] to classify image points. Here we assume that an image pixel and its label (“trail” or “not_trail”) are concurrently generated as follows. First, a hidden factor, or node, is chosen according to... |

12 |
RBSim: Geographic Simulation of Wilderness Recreation Behaviour
- GIMBLETT, RICHARDS, et al.
- 2001
(Show Context)
Citation Context ... such as GPS navigation become more widespread. Being able to monitor trails automatically will support natural resource management, directly, and through research into recreation simulation modeling =-=[1]-=-. In this paper we develop a semi-automatic method for extracting trails from aerial and satellite images. We assume that the two end points of the trail are given. Our task is to find the most likely... |

5 |
The Microsoft TerraServer
- Barclay, Eberl, et al.
- 1998
(Show Context)
Citation Context ...tabases are on the horizon. 3.1. Aerial Image data We use USGS DOQ imagery (Figure 1) because a public domain internet source is available for nearly the entire United States. Microsoft’s Terraserver =-=[6]-=- efficiently serves these images, which we are able to download, on-the-fly, using TopoFusion [7] software. Using TopoFusion we are able to seamlessly process GPS data from around the country. As poin... |

1 |
Pathway Extraction using Snakes with GPS
- Morris
- 2002
(Show Context)
Citation Context ...g increasingly plentiful. However, for our purpose, raw GPS data has a non-negligible amount of error, and thus we automatically lock the GPS data onto the nearby trail using the GPS-snakes algorithm =-=[2]-=-. Thus we can easily acquire a large quantity of aerial image data. We compute relatively simple feature vectors and then train systems to estimate the likelihood that an observed image patch is on a ... |