Anelia Angelova

Department of Computer Science, California Institute of Technology

1200 E. California Blvd. MC 136-93, Pasadena, CA, 91125, USA

Email: anelia [at] vision [.] caltech [.] edu

Automatically supervised dimensionality reduction
We propose a probabilistic framework for dimensionality reduction which can use automatically obtained noisy or ambiguous signals as supervision. The key idea is that even uncertain supervision information can be utilized to select lower dimensional representations which are more suitable for the task at hand. The method combines reasoning under uncertainty with dimensionality reduction. It is applied to vision-based terrain recognition using the rover's slip signals as supervision. The method enables working with high dimensional visual representations of terrains and learns in a fully autonomous fashion.

A. Angelova, L. Matthies, D. Helmick, P. Perona, Dimensionality Reduction Using Automatic Supervision for Vision-Based Terrain Learning , Robotics: Science and Systems (RSS), 2007

A. Angelova, EM Algorithm Updates for Dimensionality Reduction Using Automatic Supervision , Technical report, 2007

Learning slip from automatic supervision
We propose an automatic vision-based method for recognition of terrain types using the rover's slip signals as supervision. The method enables a fully autonomous learning and prediction of terrain types and the rover's mobility on the forthcoming terrain and is designed to work with ambiguous and noisy supervision signals.

A. Angelova, L. Matthies, D. Helmick, P. Perona, Learning Slip Behavior Using Automatic Mechanical Supervision , IEEE International Conference on Robotics and Automation (ICRA), 2007

Slip prediction
We consider prediction of rover slip from a distance using visual information as input. Slip measures the lack of progress of a wheeled ground robot while driving. Large amounts of slippage which can occur on certain surfaces, such as sandy slopes, will negatively affect rover mobility. Therefore, obtaining information about slip before entering a particular terrain can be very useful for better planning and avoiding terrains with large slip. The proposed method is based on learning from experience and consists of terrain type recognition and nonlinear regression modeling. After learning, slip prediction is done remotely using only the visual information as input.

A. Angelova, L. Matthies, D. Helmick, P. Perona, Learning And Prediction of Slip Using Visual Information, Journal of Field Robotics, 2007

A. Angelova, L. Matthies, D. Helmick, P. Perona, Slip Prediction Using Visual Information, Robotics: Science and Systems (RSS), 2006

D. Helmick, A. Angelova, M. Livianu, L. Matthies, Terrain Adaptive Navigation for a Mars Rover, IEEE Aerospace Conference, 2007

A. Angelova, L. Matthies, D. Helmick, G. Sibley, P. Perona, Learning to Predict Slip for Ground Robots, IEEE International Conference on Robotics and Automation (ICRA), 2006

Variable-length texture representation
We propose a texture representation in which the feature descriptor varies depending on the complexity of the task at hand or the misclassification penalty. This work enables faster terrain classification and more efficient use of the available resources or onboard sensors.

A. Angelova, L. Matthies, D. Helmick, P. Perona, Fast Terrain Classification Using Variable-Length Representation for Autonomous Navigation , IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2007

Data pruning
Contrary to the common belief that more training data yield better generalization, we show that the quality of the examples also matters and that the learning algorithm might be better off when some detrimental training examples are discarded. The question is which examples need to be eliminated, so as to improve generalization performance. We propose a general approach, called 'data pruning', to automatically identify and eliminate examples that are troublesome for learning with a given model. We apply the proposed method to a challenging dataset of faces, collected from the Web, achieving significant improvements in performance, especially for very noisy data. The proposed method is intended for very complex learning models which RANSAC cannot tackle.

A. Angelova, Y. Abu-Mostafa, P. Perona, Pruning Training Sets for Learning of Object Categories, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2005

A. Angelova, Data Pruning, Master's Thesis, Computer Science Department, California Institute of Technology, 2004

Dataset used in the experiments: 10,000 Web Faces

Terrain generation
CS 175. Topics in Geometric Modeling class project (2002). We propose to generate realistic stochastic 3D terrains using the Loop subdivision scheme. We first learn a probabilistic terrain model from real-life elevation data obtained from the National Geophysical Data Center. We applied DeBonet algorithm to learn dependencies of data at consequential resolutions (a statistical model of the wavelet coefficients of the data between each two consecutive levels of resolution is built). We use these models to generate terrains in a stochastic way. We first create a random terrain `seed' at low resolution, e.g. higher elevation where we want to grow a mountain, etc. Then the crude terrain outline is subdivided using the Loop subdivision scheme, at each level generating offsets sampled from the learned conditional distributions. Thus the terrains are not built randomly but follow the stochastic process learned from real life elevation maps. Example transition between lower resolution and higher resolution terrain representation.

Last updated: March 21, 2007.