@InProceedings{vaish04using, author = {Vaibhav Vaish and Bennett Wilburn and Neel Joshi and Marc Levoy}, title = {Using Plane + Parallax for Calibrating Dense Camera Arrays}, booktitle = {Proc. CVPR}, year = {2004} } @Book{krotkov, author = {Eric Paul Krotkov}, editor = {}, title = {Active Computer Vision by Cooperative Focus and Stereo}, publisher = {Springer-Verlag}, year = {1989}, OPTkey = {}, OPTvolume = {}, OPTnumber = {}, OPTseries = {}, OPTaddress = {}, OPTedition = {}, OPTmonth = {}, OPTnote = {}, OPTannote = {} } @InProceedings{wexler02bayesian, author = {Yoni Wexler, Andrew Fitzgibbon and Andrew Zisserman}, title = {Bayesian Estimation of Layers From Multiple Images}, booktitle = {Proc. ECCV}, year = {2002}, OPTannote = {By registering the images to the foreground, they effectively produce a series of images where the foreground is constant and the background changes. When the background is uniform, the problem becomes poorly constrained. To solve this, they formulate the problem as a maximum a posteriori estimation and introduce three new priors to regularize the it: clamping foreground and alpha values, spatial continuity of alpha values while respecting gradients, and a learned probability distribution for alpha.} } @InProceedings{apostoloff04bayesian, author = {Nicholas Apostoloff and Andrew Fitzgibbon}, title = {Bayesian Video Matting Using Learnt Image Priors}, booktitle = {Proc. CVPR}, year = {2004}, OPTannote = {They extend wexler's 02 paper to exploint spatial AND temporal smoothness.} } @InProceedings{tomasi98bilateral, author = {C. Tomasi and R. Manduchi}, title = {Bilateral Filtering for Gray and Color Images}, booktitle = {Proc. ICCV}, year = {1998}, OPTannote = {Performs domain (pixel space) and range (color space) filtering together -- hence called Bilateral filtering. Preserves edges but smooths out noise in same material regions of the image. For pixels near an edge, only those pixels (in the domain) which are similar in color to the pixel will be used for the filtering -- hene preserving the edge. This technique is similar to performing filtering in a combined domain+range space. The range (color) space can be in CIE-Lab color space, to provide for a Euclidean distance for similar colors.} } @InProceedings{swaminathan03perspective, author = {Rahul Swaminathan, Michael D. Grossberg and Shree K. Nayar}, title = {A Perspective on Distortions}, booktitle = {Proc. CVPR}, year = {2003}, OPTannote = { Provides 3 things: 1) a metric to quantify distortions in a view 2) a method to compute minimall distorted views from an MVI (Multi-Viewpoint image) and 3) an approximation to 2, which morphs an entire MVI to a quasi-single viewpoint perspective. Quantifying the caustic distortions from a view (of an MVI) is done by finding the parameters of the closest perspective camera which minimizes the disparity in corresponding points in the view and the image of the perspective camera. Estimating the view from a perspective camera assumes knowledge of the scene, so the MVI is backprojected to the scene, and a perspective camera is rendered from the backprojected scene. They also talk about modelling the perspective camera as a 3x4 matrix, so only solve for 11 parameters (linearly). In part 2, they seek to find the mininally distorted views from an MVI. In other words, they seek an "optimal" viewmap, mapping from MVI to view. To do this, they need some estimate of the scene. They model the scene as a primitive (like plane, or sphere) with a parameter vector s. s is distributed over some range according to a probability density function. They optimize over s to find the scene which (when projecting the MVI upon) produces the least distorted image. They then show synthetic images with a sphere ball inside a cube, and approximate the scene with a sphere. But this optimization is a slow procedure, so they seek to warp the MVI to a quasi-perspective (think warping a sphere MVI so the lines are straight). This is done with an angle-based distortion metric based on the actual scene point and the estimated scene point (didn't understand this section fully). Then, at every pixel of the MVI, we estimate a scene depth d which minimizes the angular distortion.} } @InProceedings{ramamoorthi02analytic, author = {Ravi Ramamoorthi}, title = {Analytic PCA Construction for Theoretical Analysis of Lighting Variability in Images of a Lambertian Object}, booktitle = {Proc. PAMI}, year = {2002}, OPTannote = {shows relationship between PCA and the spherical harmonic representation of incident illumination and reflection. In PCA, first 5 principal components explain most of the image variation. Computes the covariance (correlation?) matrix of a vector of images under different illumionation and shows that the a related matrix (orthoginality matrix for spherical harmonics) has eigenvectors which correspond to the principal components of the image (from PCA). For pixels distributed over the hemisphere (ie: from a single viewpoint and not from all directions from a surface), he shows that the first 5 principal components suffice (cover 95 percent of data). By shifting by the mean (as in PCA), this removes the "DC" component of the spherical harmonics.} } @InProceedings{baker98theory, author = {Simon Baker and Shree K. Nayar}, title = {A Theory of Catadioptric Image Formation}, booktitle = {Proc. ICCV}, year = {1998}, OPTannote = {catadioptric sensor == combination lens and mirrors to capture a wider FOV. argue that cata. sensor should form a single effective viewpoint. paper does 3 things: 1) dervies complete class of single-lens single-mirror and single effective viewpoint (which are conic sections) 2) computes spatial resolution of cata. sensor in terms of camera resolution (defined as diff Area/diff solid angle) 3) analysis of defocus blur due to use of a curved mirror} } @InProceedings{boykov99fast, author = {Yuri Boykov and Olga Veksler and Ramin Zabih}, title = {Fast Approximate Energy Minimization via Graph Cuts}, booktitle = {Proc. ICCV}, year = {1999}, OPTannote = {A minimization technique for energy functions of the form Edata + Esmooth. The energy function takes in a mapping. A mapping is a function from pixels to labels (ie: pixels to disparities, pixels to colors, etc.). Edata represents difference between observed label and the assigned label. Esmooth ensures some continuity/smoothness in the labeling. Note: assumes that the E function is smooth, which may not be the case a object boundaries, shadows, depth discontinuities, etc. Energy functions can be metric (preseves triangle inequality) or semi-metric. The two types of minimizing moves are alpha-beta swaps (swaps pixels of either label alpha or beta), or alpha-expansions (expands a label to more pixels). These methods produce local minimums even when large moves are allowed. Basic minimization algorithm (alpha-beta swap): pick 2 labels, find the optimal number of pixels to swap labels, compute the new energy function, if the new energy function is lower than previous, set the new one, and continue with new pair of labels. If the energy is higher, than we're done (at local min). Even if local min is found, they state (didn't prove in paper) that the local min is within some bound of the global minimum. To find the optimal number of pixels to swap (and to which label), they use graph cuts. They map a particular labeling (and alpha-beta) to a graph, and show that a graph cut corresponds to a new labeling where the new label is 1 alpha-beta swap from the old labeling. By assigning certain weights to each edge, they show that the weight of the graph cut is exactly the energy of that labeling. Then, to find the labeling with smallest energy, one needs to find the min graph cut. The same procedure is presented for alpha-expansions, but now an auxillary node is also presented -- to disambiguiate cases where 2 pixels have different labels during an alpha expansion, and when they have the same labels during an alpha expansion. Their results don't really show how good/bad their algorithm is -- they compare their work with simulated annealing and normalized correlation, but who's to say that one is better than the other? Vaibhav stated a good point: vision people don't do segmentation for the sake of segmentation -- they use it for something. This paper compares its algorithm with other segmentations, but don't show whether it's useful for the actual application. Another question is: what types of energy fuctions is this good for? There's another paper on this -- need to get reference from Vaibhav.} } @TechReport{szeliski1993Recovering3D, author = {Richard Szeliski and Sing Bing Kang}, title = {Recovering 3D Shape and Motion from Image Streams using Non-Linear Least Squares}, institution = {Digital Equipment Corporation, Cambridge Research Lab}, year = {1993}, OPTannote = {Non-linear least squares to solve for 3d structure, motion and camera parameters simultaneously, smart parameterization of unknowns and formulation of Hessian and gradient vectors.} } @InProceedings{faugeras98, author = {Olivier Faugeras and Renaud Keriven}, title = {Complete Dense Stereovision using Level Set Methods}, booktitle = {Proc. ECCV}, year = {1998}, OPTannote = {scene is represented by a level set, and level sets are evolved to find the final scene function. A PDE describes how level sets go from one to another, the metric is a correlation that is correct to an affine transformation.} } @TechReport{kiriakos98, author = {Kiriakos N. Kutulakos and Steven M. Seitz}, title = {A Theory of Shape by Space Carving}, institution = {U. Rochester C.S. Dept}, year = {1998}, number = {TR #692}, month = {May}, OPTannote = {scene is represented by voxels, initially a rough volume is provided, then voxels which are not "photo-consistent" are removed. "Photo-consistent" voxels are voxels that are consistent with a given lighting model (ie: lambertian, etc.)} } @Book{trucco98, author = {Emanuele Trucco and Alessandro Verri}, ALTeditor = {}, title = {Introductory Techniques for 3-D Computer Vision}, publisher = {Prentice-Hall}, year = {1998}, } @Book{hartley2000, author = {Richard Hartley and Andrew Zisserman}, ALTeditor = {}, title = {Multiple View Geometry in computer vision}, publisher = {Press Syndicate of the University of Cambridge}, year = {2000}, }