CISC 849: Advanced Topics in Computer Applications - Spring 2010: Computer Vision

Student Projects

3D Motion Estimation Using Manifold Learning

by Rohith Mv

Goals:

Estimate non-rigid motion between surfaces corresponding to human face
Optimize the estimation by learning motion characteristics from training data

Summary: The problem of 3D registration is ill-posed and requires some nature of regularization to render it solvable. Generic smoothness constraints may not always be sufficient to yield desired results. We propose a series of regularization schemes based on manifold learning for estimating the non-rigid deformation between surfaces. We explore the motion estimation of meshes corresponding to human faces with various expressions. We propose a novel method for regularization using representation of non-rigid deformation fields as Bézier surfaces. The performance is improved by reducing the dimension of the deformation space through manifold learning. A quantitative comparison of various linear and non-linear dimensionality reduction techniques is provided. We also present a method to characterize the deformations in different parts of the model using segmentation of the deformation fields. It is applied to a situation where different face models are deformed with various expressions and the forces corresponding to deformations in regions of the face are characterized. We show using real 3D face data that the characteristics learnt may be used in segmenting the face.

Face Expression Recognition From 3D Models

by Gowri Somanath

Goals:

Given 3D mesh model for neutral and an expression for a person, classify the expression into one of the trained categories.

Methods:

3D data capture using stereo, triangulation and mesh creation
Curvature and surface type calculation
Dividing model into regions for feature extraction
Training a random forest (RF) on the features
Testing and accuracy assessment

Land Cover Change Detection Analysis

by Feray Demirci, Caghan Demirci

Goals:

Register two satellite images of the same area, taken at different times, using intensity-based registration
Detect the different land cover classes in the images
Analyze the change in a land cover class between the two images, which gives us this land cover change over a certain time period

Methods:

Intensity-based registration, using similarity transformation, SSD, gradient (steepest) descent and pyramid scheme
Dimensionality reduction using PCA
k-means clustering on the first principal components, after determining optimal k (number of clusters)
Change detection using image differencing

Dense Stereo for Fisheye Cameras

by Yan Lu

Goals:

Detect obstacles
Robot navigation

Summary: In this report, work of reconstructing 3D scene from stereo fisheye cameras is presented. The two fisheye cameras are mounted on a ground vehicle. The goal of this work is to detect obstacles, both positive and negative, through 3D reconstruction, which can be used as a cue for robot navigation. Basically, the procedures of realizing such a system are: (1) undistort and rectify images, (2) calculate disparity, (3) reproject image to 3D, and (4) detect and represent obstacles. The system is tested and validated on image sequences that have trails with various surroundings, such as grass, bushes, and trees. In addition, a utility OpenGL program is written in order to visualize the reconstructed 3D scene.

3D Rigid Motion Estimation

by Divya Muppaneni, Nimmy Kurien, Rag Mayur Chevuri

Goals:

To estimate the 3D motion of rigid objects using the algorithms: Small Angle Approximation and Scatter Matrix

Methods:

Conducting experiments on synthetic data and validating results
Using 3D face data and implementing these algorithms to estimate rotations
Comparing elapsed time for both algorithms

Big View Video/Image Modeling and Synthesis Lab Home
VIMS Home People Research Publications Workshops Achievements Courses Software Data Photos login	CISC 849: Advanced Topics in Computer Applications - Spring 2010: Computer Vision Student Projects 3D Motion Estimation Using Manifold Learning by Rohith Mv Goals: Estimate non-rigid motion between surfaces corresponding to human face Optimize the estimation by learning motion characteristics from training data Summary: The problem of 3D registration is ill-posed and requires some nature of regularization to render it solvable. Generic smoothness constraints may not always be sufficient to yield desired results. We propose a series of regularization schemes based on manifold learning for estimating the non-rigid deformation between surfaces. We explore the motion estimation of meshes corresponding to human faces with various expressions. We propose a novel method for regularization using representation of non-rigid deformation fields as Bézier surfaces. The performance is improved by reducing the dimension of the deformation space through manifold learning. A quantitative comparison of various linear and non-linear dimensionality reduction techniques is provided. We also present a method to characterize the deformations in different parts of the model using segmentation of the deformation fields. It is applied to a situation where different face models are deformed with various expressions and the forces corresponding to deformations in regions of the face are characterized. We show using real 3D face data that the characteristics learnt may be used in segmenting the face. Face Expression Recognition From 3D Models by Gowri Somanath Goals: Given 3D mesh model for neutral and an expression for a person, classify the expression into one of the trained categories. Methods: 3D data capture using stereo, triangulation and mesh creation Curvature and surface type calculation Dividing model into regions for feature extraction Training a random forest (RF) on the features Testing and accuracy assessment Land Cover Change Detection Analysis by Feray Demirci, Caghan Demirci Goals: Register two satellite images of the same area, taken at different times, using intensity-based registration Detect the different land cover classes in the images Analyze the change in a land cover class between the two images, which gives us this land cover change over a certain time period Methods: Intensity-based registration, using similarity transformation, SSD, gradient (steepest) descent and pyramid scheme Dimensionality reduction using PCA k-means clustering on the first principal components, after determining optimal k (number of clusters) Change detection using image differencing Dense Stereo for Fisheye Cameras by Yan Lu Goals: Detect obstacles Robot navigation Summary: In this report, work of reconstructing 3D scene from stereo fisheye cameras is presented. The two fisheye cameras are mounted on a ground vehicle. The goal of this work is to detect obstacles, both positive and negative, through 3D reconstruction, which can be used as a cue for robot navigation. Basically, the procedures of realizing such a system are: (1) undistort and rectify images, (2) calculate disparity, (3) reproject image to 3D, and (4) detect and represent obstacles. The system is tested and validated on image sequences that have trails with various surroundings, such as grass, bushes, and trees. In addition, a utility OpenGL program is written in order to visualize the reconstructed 3D scene. 3D Rigid Motion Estimation by Divya Muppaneni, Nimmy Kurien, Rag Mayur Chevuri Goals: To estimate the 3D motion of rigid objects using the algorithms: Small Angle Approximation and Scatter Matrix Methods: Conducting experiments on synthetic data and validating results Using 3D face data and implementing these algorithms to estimate rotations Comparing elapsed time for both algorithms Page last modified on August 11, 2010, at 03:36 AM