Computer Vision Misc Reading Group

1. Spectral Histogram Based Face Detection by Christopher Waring and Xiuwen Liu. Proceedings of the International Joint Conference on Neural Networks, Volume: 2, July 20-24, 2003. Paper here.

2. Rotation Invariant Face Detection Using Spectral Histogram and SVMs. CVPR submission.

A technical report can be found here.

Extreme Components Analysis
Max Welling, Felix Agakov, Christopher K. I. Williams. Paper here.

and if I have time:

Non-linear CCA and PCA by Alignment of Local Models, Paper here.

Enjoy the reading.

The Unified Propagation and Scaling Algorithm, Y. W. Teh and M. Welling, NIPS 2001. Paper here.

The paper develops a unified framework for the above manifold learning algorithms and extensions for out-of-sample data. A more detailed tech report is Spectral Clustering and Kernel PCA are Learning Eigenfunctions. Paper here.

I'll be talking about Unsupervised Improvement of Visual Detectors using Co- Training by A Levin, P Viola, and Y Freund, from ICCV 2003. You can find the paper here or here.

Bags of Pixels is a technique for representing images as unordered sets of tuples. Finding optimal orderings for these tuples across sets of images automatically estimates pixel correspondences over the images and compacts the image representations into a subspace conducive to PCA. Furthermore, resultant PCA basis vectors have visual characteristics that give rise to compelling intuitive interpretations. This technique can also be supplied to other sampled data, like audio recordings.

For more background, you may also wish to take a look at this paper, which describes the convex optimization technique of the Bags of Pixels paper in slightly greater detail.

And if I have time, I may also talk about Wei & Levoy's Siggraph 2000 paper on texture synthesis using a non-parametric model:
Fast Texture Synthesis using Tree-structured Vector Quantization. Paper here.

Robert E. Schapire.
The boosting approach to machine learning: An overview.
In MSRI Workshop on Nonlinear Estimation and Classification, 2002.

Jerome Friedman, Trevor Hastie and Robert Tibshirani.
Additive logistic regression: a statistical view of boosting.
The Annals of Statistics, 38(2):337-374, April, 2000.

Robert E. Schapire and Yoram Singer.
Improved boosting algorithms using confidence-rated predictions.
Machine Learning, 37(3):297-336, 1999.

1. Automatically generating 3D models of low-contrast brain structures from volumetric images.
2. Using MCMC-type sampling techniques to evaluate sources of error in registration/segmentation algorithms.
3. Selecting shape features from 3D anatomical structures for predictive statistical models. In other words, building statistical models that tell you things like, "probability that you will get Alzheimer's at time T, given that your hippocampus looks like X"

I will also touch on a couple of feature-based registration projects that I am helping to give advise to.

Texture Synthesis

In our 1999 work ([Efros and Leung,'99]) we tried to address this shortcoming by proposing a very different and extremely simple way of synthesizing textures locally, one-pixel-at-a-time. In recent years this algorithm has been used for synthesiszing a large spectrum of textures as well as filling holes in textured regions. In our latest work ([Efros and Freeman,'01]) we show that similarly good results can be obtained by an even simpler (and much faster) procedure of quilting together patches of the input texutre. Moreover, we demonstrate a method of transfering texture from one object onto another (e.g. rendering a man's face with rice).

The two papers are available here and here.

I'll try to present so that no reading is required beforehand. If you can't wait, there is a brief discussion of 6 DOF motion from image and inertial measurements here and a much more detailed discussion, which includes the Hyperion results, is available here.

There's no paper describing the smalls tracker at this point.

Selective Sampling With Redundant Views
Ion Muslea, Steven Minton, and Craig A. Knoblock

link to paper

And

"Mutual Boosting for Contextual Inference" by Michael Fink and Pietro Perona from NIPS 2003 Paper here

The main links between the papers are the obvious: boosting and multiclass detection. The main reason that I've listed both of them is because I don't think either one will fill the entire meeting. But then again, you never know.

(1) "Interpreting and Extending Classical Agglomerative Clustering Algorithms using a Model-Based Approach". In ICML, 2002 Link here , and
(2) "Multiobjective Data Clustering", M. Law, A. Topchy, A.K. Jain, CVPR 2004. Link here

The first shows how common heuristic clustering algorithms are each equivalent to a hierarchical model-based method, and offers implementers intuition on how to modify and choose between algorithms / distance measures.

The second presents an approach that uses multiple clustering objective functions simultaneously, and tries to find a robust, least conflicting partition of the dataset to be clustered.

Weakly Supervised Learning of Visual Models and Its Application to Content-Based Retrieval
by Cordelia Schmid, IJCV 56(1/2), 7-16, Jan/Feb 2004.

In this paper, Schmid trains visual models for image retrieval by building up libraries of significant rotation-invariant descriptors from positive and negative class examples. Aside from this binary labeling, no other handholding is performed.

A local link to the paper can be found here.

BoostMap: A Method for Efficient Approximate Similarity Rankings
Vassilis Athitsos, Jonathan Alon, Stan Sclaroff, and George Kollios
Link here

Depending on the time and demand, I might go on to talk about this one by one of the Locally Linear Embedding guys. It won best paper:

Unsupervised Learning of Image Manifolds by Semidefinite Programming
Kilian Q. Weinberger and Lawrence K. Saul
Link here

This paper was presented at the POCV workshop of CVPR 2004, on June 28 morning. The top-down approach uses prior knowledge about an object to guide the segmentation, whereas the bottom-up approach first segments the image into regions and then groups the object-related regions together. This paper shows how to combine these two approaches together.

Enjoy reading!!

I'll also talk about on-going work in applying similar techniques to music retrieval.

The authors encode contextual information using learnt label pattern fields and combine it with low-level classification using CRFs.

Here is the abstract:

Replicability of machine learning experiments measures how likely it is that the outcome of one experiment is repeated when performed with a different randomization of the data. In this paper, we present an estimator of replicability of an experiment that is efficient. More precisely, the estimator is unbiased and has lowest variance in the class of estimators formed by a linear combination of outcomes of experiments on a given data set.We gathered empirical data for comparing experiments consisting of different sampling schemes and hypothesis tests. Both factors are shown to have an impact on replicability of experiments. The data suggests that sign tests should not be used due to low replicability. Ranked sum tests show better performance, but the combination of a sorted runs sampling scheme with a t-test gives the most desirable performance judged on Type I and II error and replicability.

This paper describes a new way of making hyperspectral images free of motion distortion by combining diffraction gratings with a CT-style inverse problem formulation. I'll describe this new sensor and outline the computations involved.

The paper was scanned in, so the PDF file is quite large (14MB).
As an alternative, it is also available in OpenOffice format (.SXW) or in MS Word format (.DOC), each about 2MB.

Statistical Context Priming for Object Detection, Torralba, et. al 2001
Multiscale Conditional Random Fields for Image Labeling, He et. al 2004
Mutual Boosting for Contextual Inference, Fink, et. al, 2003

This paper describes an exact, low-polynomial time solution to finding a MAP labeling for two-label, N-dimensional Markov Random Fields. This is used for interactive foreground/background extraction, where the labeling of image pixels simply determines whether an object is foreground or background. It has similarly been used for texture synthesis. After discussing this paper we'll talk about its extensions to multi-class problems, and then we'll look at various direct extensions to this paper at SIGGRAPH this year.

Available here, and also here.

Max-Margin Markov Networks, B. Taskar, C. Guestrin and D. Koller. Neural Information Processing Systems Conference (NIPS03), Vancouver, Canada, December 2003.

The paper can be found here.

This work is an extension of SVM to structured or relational data (e.g. chains or MRFs).

The paper describes LO-RANSAC 3-LAF – a new algorithm for the correspondence problem; Exploiting processes proposed for computation of affineinvariant local frames, three point-to-point correspondences are found for each region-to-region correspondence. Consequently, it is sufficient to select only triplets of region correspondences in the hypothesis stage of epipolar geometry estimation by RANSAC.

The paper can be found here.

The following papers are all related to this talk:

Abstract:
We consider the problem of recovering an underwater image distorted by surface waves. A large amount of video data of the distorted image is acquired and the problem is posed in terms of understanding the statistics of local patches in the image plane. This challenging reconstruction task can be formulated as a manifold learning problem, such that the center of the manifold is the image of the undistorted patch. To compute the center, we present a new technique to estimate global distances on the manifold. Our technique achieves robustness through convex flow computations and solves the "leakage" problem inherent in recent manifold embedding techniques.

Joint work with Volkan Isler, Jianbo Shi and Mirko Visontai.

Accepted to NIPS'04, draft can be found here.

Programmable Imaging using a Digital Micromirror Array
S. K. Nayar, V. Branzoi, and T. Boult
Proceedings of IEEE Conference on Computer Vision and Pattern Recognition
Washington DC, June 2004.

The paper can be found here.

I will also try to discuss the different sensors built at Columbia on dynamic range.

Abstract:
We introduce a new class of distinguished regions based on detecting the most salient convex local arrangements of contours in the image. The regions are used in a similar way to the local interest points extracted from gray-level images, but they capture shape rather than texture. Local convexity is characterized by measuring the extent to which the detected image contours support circle or arc-like local structures at each position and scale in the image. Our saliency measure combines two cost functions defined on the tangential edges near the circle: a tangential-gradient energy term, and an entropy term that ensures local support from a wide range of angular positions around the circle. The detected regions are invariant to scale changes and rotations, and robust against clutter, occlusions and spurious edge detections. Experimental results show very good performance for both shape matching and recognition of object categories.

The paper can be found here. (If that link doesn't work for some reason, you should be able to search for the paper on IEEE Xplore.)

Paper here.

This paper is about how to use semidefinite programming to learn the kernel matrix that can "unfold" the manifold when it maps the input into the feature space. (The input high-dimensional data is assumed to lie in some low-dimensional manifold).