Year 2002 (in reverse chronological order)
Date  Presenter  Description 
12/4/2002
Wednesday 
Raghu Rao 
Please note the room change for this meeting  we will be meeting in NSH 4513!
I will be talking about the following ECCV02 paper: Multiscale EMICP: A Fast and Robust Approach for Surface Registration S. Granger and X. Pennec. Paper here. In this paper, they formulate the problem of rigid registration of 3D surfaces as a maximum likelihood estimation of transformation and matches. They then proceed to solve this problem efficiently using EM principles. They show that in the case of gaussian noise, this new formulation approximately corresponds to the usual ICP. They claim spectacular improvements over the standard ICP. 
11/27/2002

No Meeting.  Thanksgiving. 
11/20/02
Wednesday 
Goksel Dedeoglu 
I would like to talk about the following ICCV '01 paper:
Z. Ying and D. Castanon, "Feature Based Object Recognition using Statistical Occlusion Models with Onetoone Correspondence". Paper here. In addition, Google located two presentation materials from their lab that may be better suited for a quick overview of the approach. These are not 100% about their ICCV paper though: ONR/GTRI Workshop on Target Tracking and Sensor Fusion, June 2002. Slides here. 
11/13/2002
Wednesday 
Sanjiv Kumar 
Continuing the earlier discussion on Tree Structured Belief Networks
(TSBN), I will present the work on Dynamic
Trees by Nick Adams and Chris Williams. The main papers I will present
are:
DTs: Dynamic Trees, by C. K. I. Williams and Nicholas J. Adams, NIPS 1999. Dynamic Trees: Learning to Model Outdoor Scenes, by Nicholas J. Adams and Christopher K. I. Williams, ECCV 2002. Other than the above papers, there are variants of dynamic trees, such as Mean Field Dynamic Trees, Sparse Dynamic Trees and Dynamic Positional Trees. I have kept a tar file including five papers on DTs at the following link. All the papers on DTs are available as separate files at the following link. 
10/30/2002
Wednesday 
Bob Wang 
I will talk about Andrew Johnson and Roberto Manduchi's
surface integration paper in 3DPVT'02 according to Martial's suggestion:
"Probabilistic 3D Data Fusion for Adaptive Resolution Surface Generation", 3D Data Processing Visualization and Transmission, 2002. This paper can be downloaded here. 
10/23/2002
Wednesday 
Tom Minka  I will talk about the local linear embedding work of Roweis and Saul. 
10/16/2002
Wednesday 
Chuck Rosenberg 
I will discuss the Viola and Jones face detector papers.
This paper provides a good concise description of the method: Rapid Object Detection using a Boosted Cascade of Simple Features. Paul Viola and Michael Jones. CVPR 2001. Paper This paper is 25 pages long and has more detail: Robust Realtime Object Detection Paul Viola and Michael Jones. Presented at ICCV 01 workshop on statistical and computational theories of vision. Also submitted to IJCV. Paper. This paper describes some additional work on a small tweak to AdaBoost to penalize false negatives: Fast and Robust Classification using Asymmetric AdaBoost and a Detector Cascade. Paul Viola and Michael Jones. NIPS 2001. Paper. 
10/9/2002
Wednesday 
Henry Schneiderman 
I'll be talking about some recent work in classifier design.
The title of my talk will be "Using Competition to Design a Modified Bayes Classifier". Abstract: A modified naive Bayes classifier represents the joint statistics of small subsets of variables while treating the subsets as statistically independent. This talk introduces an automatic training method that decomposes the variables into such subsets. The training method generates a large set of candidate subsets and trains "subclassifiers" over each subset. An empiricallybased competition selects a combination of these subclassifiers to form the overall classifier. Using this method, I have constructed object detectors for human faces and telephones. 
10/2/2002
Wednesday 
Cristina Dima 
I will talk about the following paper:
"Soft Margins for AdaBoost" by G. Ratsch, T. Onoda, K.R. Muller, Machine Learning 2000. The paper can be downloaded here. Here is the abstract: "Recently ensemble methods like AdaBoost have been applied successfully in many problems, while seemingly defying the problems of overfitting. AdaBoost rarely overfits in the low noise regime, however, we show that it clearly does so for higher noise levels. Central to the understanding of this fact is the margin distribution. AdaBoost can be viewed as a constraint gradient descent in an error function with respect to the margin. We find that AdaBoost asymptotically achieves a hard margin distribution, i.e. the algorithm concentrates its resources on a few hardtolearn patterns that are interestingly very similar to Support Vectors. A hard margin is clearly a suboptimal strategy in the noisy case, and regularization, in our case a ``mistrust'' in the data, must be introduced in the algorithm to alleviate the distortions that single difficult patterns (e.g. outliers) can cause to the margin distribution. We propose several regularization methods and generalizations of the original AdaBoost algorithm to achieve a soft margin. In particular we suggest (1) regularized AdaBoostReg where the gradient decent is done directly with respect to the soft margin and (2) regularized linear and quadratic programming (LP/QP) AdaBoost, where the soft margin is attained by introducing slack variables. Extensive simulations demonstrate that the proposed regularized AdaBoosttype algorithms are useful and yield competitive results for noisy data." 
9/25/2002
Wednesday 
Sanjiv Kumar 
I will be presenting a PAMI paper dealing with
the Tree Structured Bayesian Networks (TSBNs) applied to the problem of
image region classification. The reference of the paper is:
Xiaojuan Feng, Williams, C.K.I., and Felderhof, S.N., "Combining belief networks and neural networks for scene segmentation", PAMI, April, 2002, pp. 467483. A copy of the paper can be obtained from the following link. 
9/18/2002
Wednesday 
Owen Carmichael
and Chuck Rosenberg 
BMVC 2002 Overview 
9/11/2002
Wednesday 
Hannes Kruppa 
Towards object detection using multiple background models
This is a wrapup of the research I did on using multiple background models for object detection. I focused on face detection since faces naturally do occur within arbitrary backgrounds. The hope is (was? ;) that background estimation will allow to improve detection performance. I will describe the specific approach I took and the results I got so far 
9/4/2002
Wednesday 
Bogdan Matei 
Georegistration
The visitor on Wed. is Bogdan Matei, now at Sarnoff, and he will talk about georegistration. No abstract or paper, we'll just have to be there :) 
8/28/2002
Wednesday 
Owen Carmichael 
I'll be giving a practice version
of a 20minute conference talk about object recognition from edge cues,
which covers the material in our BMVC paper, Object Recognition by a Cascade of Edge Probes. The paper is
here.
This will also be an administrative meeting, so please be sure to attend. 
8/9/2002
Friday 
Shyjan Mahamud  Discriminative Distance Measures for Object Detection 
7/29/2002
Monday 
Peng Chang  Robust Tracking and Structure from Motion with Sampling Method 
7/23/2002
Tuesday 
Daniel Huber  Automatic ThreeDimensional Modeling from Reality 
7/16/2002
Tuesday 
Matt Deans  Bearings Only Localization and Mapping 
7/10/2002
Wednesday 
Martial Hebert 
I will try to attempt to begin talking about
K. Mikolajczyk and C. Schmid's paper in ECCV02. The paper is
here.
For the extremely courageous, 2 (long) background papers are: widbas.pdf: Different approach by Van Gool's group. Earlier version was presented by Bart a year or so ago. affine.pdf: Background used in the Schmid paper for affine invariance and scale selection by Lindeberg. For the courageous.... The context of the paper is in wide baseline stereo, but it does have potentially interesting applications for recognition/detection, etc. 
7/3/2002
Wednesday 
Owen Carmichael 
I'll talk at the reading group about a paper Tom pointed me
to, namely
Prosection Views: Dimensional Inference through Sections and Projections," G. W. Furnas, A. Buja, Journal of Computational and Graphical Statistics 3, 323385 (1994). I couldn't find an electronic version where all the figures were viewable. The closest is here, which only has a couple of broken figures. I put 10 paper copies on top of the cabinet just inside the 4th floor entrance to RI (i.e. the usual spot). If all goes well during the meeting you should understand the content of the paper without having read it ahead of time. The gist of the paper is that if you have a highdimensional data set, you can infer some things about the intrinsic dimensionality of the data by taking random projections and random slices of the data until you end up with something you can look at in 2D. This is an alternative to strictly numerical approaches to dimensionality inference, like PCA. 
6/26/2002
Wednesday 
Hannes Kruppa 
Hannes is a PhD student of Bernt Schiele at ETH Zurich
(www.vision.ethz.ch/pccv) visiting the Robot Learning Lab
of Sebastian Thrun for this summer.
The title of his talk is: "Towards Contextual Models of Appearance For Object Detection" I will present a snapshot of my own current research on multiview face detection in arbitrary still images. Note that this is work in progress at a fairly early stage. The main idea is to model the dependencies between object appearance and the global scene as well as other contextual information, hoping that this will help in the detection task. The talk starts with a theoretical and empirical analysis of the failure modes of Henry Schneiderman's detector, which is used as a starting point for this research. Then, specific context components are proposed that could help to overcome such failures. Some preliminary results are presented here. Finally, a specific representation for contextual appearance models is proposed. The talk concludes with a list of my "most pressing, open questions" hoping to stir discussion. 
6/17/2002
Monday 
Raghu Rao 
I will be discussing "Tensor Voting" by Gerard
Medioni and group. You can find many papers online, but I found this one
easier to understand.
C.K. Tang, M.S. Lee, G. Medioni, "Tensor Voting", in Recent Advances in Perceptual Organization, Kluwer. Paper. 
6/10/2002
Monday 
Sanjiv Kumar  ECCV Review Session 
6/3/2002
Monday 
Nicolas Vandapel  Surface Matching by 3D Point's Fingerprint. Y. Sun and M.A. Abidi. ICCV'01. Paper. 
5/20/2002
Monday 
Goksel Dedeoglu 
The paper for Monday, 05/20:
Shape Matching and Object Recognition Using Shape Contexts Serge Belongie, Jitendra Malik and Jan Puzicha PAMI, 24(4):509522, April 2002. Paper. 
5/13/2002
Monday 
David Tolliver 
I'll present:
"Robust, online appearance models for vision tracking.", Jepson, A.D., Fleet, D.J. and ElMaraghi, T. CVPR 01. Runnerup for best paper. PSGZ Paper. PDF Paper. (The quality of the PDF file is quite poor.) 
5/6/2002
Monday 
Tom Minka 
I will discuss the work of Peter Meer and his students on nonlinear
errorsinvariables estimation, a technique which has broad applicability
in computer vision in addition to being theoretically interesting.
The canonical example is fitting a curved surface to scattered points,
which arises for example in learning a deformable template.
The most recent paper is: "Registration via direct methods: A statistical approach" J. Bride and P. Meer, CVPR'01. Paper.

4/29/2002
Monday 
Charles Rosenberg 
I will discuss the work of Brendan Frey and Nebojsa Jojic on
incorporating latent variables to model clutter and transformations
when modeling image data using Gaussians, mixtures of Gaussians,
component analyzers or mixtures of component analyzers.
They have published quite a few papers on this topic. The paper that I will discuss is a preprint of a paper that was submitted to PAMI, but does not seem to have been published yet: Brendan J. Frey and Nebojsa Jojic 2000. Transformationinvariant clustering and dimensionality reduction. To appear in IEEE Transactions on Pattern Analysis and Machine Intelligence, Submitted Nov. 2000. Paper.

4/22/2002
Monday 
Sanjiv Kumar 
I will discuss the remaining part of my earlier
talk on Relevance Vector Machine (RVM). The main paper of discussion will
be:
M. A. Figueiredo, A. K. Jain, "Bayesian Learning of Sparse Classifiers," IEEE CVPR, Hawaii 2001. Paper. I will also give a quick review of the RVM paper, which we have already discussed in my last talk, for the purpose of continuity. The paper on RVM is: M. E. Tipping, "The Relevance Vector Machine," NIPS, San Mateo, 2000. Paper. 
4/15/2002
Monday 
Everyone  Short scheduling meeting. 
4/1/2002
Monday 
Matt Deans 
Y. Caspi and M. Irani, Alignment of NonOverlapping Sequences. IEEE
International Conference on Computer Vision (ICCV), Vancouver, July
2001.
Paper.
This paper received the Honorable Mention for the 2001 Marr Prize. Two important but optional background papers: Tracking from Multiple View Points: Selfcalibration of Space and Time, in DARPA Image Understanding Workshop (1998). Paper. Y. Caspi and M. Irani, A step Towards SequencetoSequence Alignment. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2000. Paper. 
3/25/2002
Monday 
Owen Carmichael 
I'll give an overview of what illumination cones are and what we know
about them, so you don't need to read any papers if you don't want.
However I'll be drawing from these 3 papers on the topic which you can
grab if you like:
P. Belhumeur and D. Kriegman, "What Is the Set of Images of an Object Under All Possible Illumination Conditions?" Int. Journal of Computer Vision, 28(3), 1998, PP. 24560. Paper. A conference version of this paper won the best paper award at CVPR '96. It gives a derivation of what illumination cones are and how you can estimate them from 3 example images. P. Belhumeur, A. Georghiades and D. Kriegman, "From Few to Many: Illumination Cone Models for Face Recognition Under Variable Lighting and Pose," IEEE Trans. PAMI, 2001, pp.643660. Paper. An illumination cone only covers views of the object at one pose. So why not many different cones for many different poses. NinePoints of Light: Acquiring Subspaces for Face Recognition under Variable Lighting. KuangChih Lee, Jeffery Ho, David Kriegman / IEEE Conf. on Computer Vision and Pattern Recognition, 2001, pp.519526. Paper. Shows that if you set up 9 light sources and a camera, and snap 9 photos, you can use those training images to do recognition over a wide range of lighting changes. This paper makes a lot more sense if you also read this one: Lambertian Reflectance and Linear Subspaces. Ronen Basri and David Jacobs, ICCV 2001. Paper. 
3/18/2002
Monday 
Everyone 
CVPR / ICCV Paper Discussion
Conference sweep type meeting. The List of Papers and Ideas for 2002 compiled from our group meeting by Daniel Huber. 
3/11/2002
Monday 
Cristian Dima 
On combining classifiers, by Kittler, Hatef, Duin and Matas, from
TPAMI March 1998
A theoretical study on six classifier fusion strategies, by Ludmila Kuncheva, from TPAMI Feb 2002 Paper. Time permitting: Bayesian Fusion of Color and Texture Segmentations, by Roberto Manduchi, from ICCV 99. Paper. 
3/4/2002
Monday 
Tom Minka 
I will give an overview of the following paper. I will show that
the statistical explanations given in the paper are flawed, but that their
methods do make sense if you assume the right probabilistic image model.
This model has connections with eigenvectorbased segmentation and matching
by mutual information.
"Similarity templates for detection and recognition" Chris Stauffer and Eric Grimson, CVPR'01. Paper. 
2/25/2002
Monday 
Goksel Dedeoglu  "A new algorithm for nonrigid point matching", H. Chui and A. Rangarajan, IEEE Conference on Computer Vision and Pattern Recognition 2000, volume 2, pages 4451. Paper. Web Page. Online Demos. 
2/18/2002
Monday 
Dennis Strelow 
I'll try to give an overview on recent work on integrating
visual and inertial sensors for camera motion estimation.
I will present so that you don't have to read the papers, but FYI here are the citations from the two relevant papers from CVPR 2001: SangHack Jung and Camillo J. Taylor. Camera trajectory estimation using inertial sensor measurements and structure from motion results. CVPR '01. Vol. 2, p. 732737. Paper. Andrew S. Davison and Nobuyuki Kita. 3D Simultaneous Localisation and MapBuilding Using Active Vision for a Robot Moving on Undulating Terrain. CVPR '01, Vol 1, p. 384391. Paper. If time allows I will include an overview of the following related, but slightly less recent work: Clark F. Olson, Larry H. Matthies, Marcel Schoppers, and Mark W. Maimone. Robust stereo egomotion for long distance navigation. CVPR '00. 453458. Paper. Gang Qian, R. Chellappa, and Qinfen Zheng. Robust structure from motion using inertial data. Accepted by JOSA. Paper. Andreas Huster and Stephen M. Rock. Relative position estimation for interventioncapable AUVs by fusing vision and inertial measurements. In Proceedings of the 12th International Symposium on Unmanned Untethered Submersible Technology, Durham, NH, August 2001. Autonomous Undersea Systems Institute. Paper. 
2/11/2002
Monday 
Sanjiv Kumar 
I will be speaking about the 'Bayesian Learning of Sparse Classifiers'.
I will also discuss the
related concept of Relevance Vector Machine (RVM) and its advantages (and
disadvantages) over the Support Vector Machine (SVM). The references are:
M. A. Figueiredo, A. K. Jain, "Bayesian Learning of Sparse Classifiers," IEEE CVPR, Hawaii 2001. Paper. M. E. Tipping, "The Relevance Vector Machine," NIPS, San Mateo, 2000. Paper. 
2/4/2002
Monday 
Owen Carmichael 
I'll be talking about the tensor rank decomposition and two
papers that have applied it in computer vision. The tensor rank
decomposition is yet another way of approximating a set of
highdimensional data with a small number of basis vectors, so it's
comparable to PCA, ICA, etc.
I'll describe what t.r.d. is and present this paper, Linear Image Coding For Regression and Classification Using The TensorRank Principle, A. Shashua and A. Levin, CVPR '01, pg I42, Paper, which applies t.r.d. to image representation and classification. I'll also talk about the first 2 sections of Efficient Deformable Filter Banks, Roberto Manduchi, Pietro Perona and Doug Shy. IEEE Trans. on Signal Processing. Vol. 46, N. 4, Pag. 11681173. 1998, Paper, where they want to approximate a set of 2D convolution filters by sets of 1D steerable filters. 
2/1/2002
Friday 
Charles Rosenberg 
I will be continuing my overview of papers from the recent NIPS conference
which utilize unlabeled or partially labeled training data.
The primary paper which I will be discuss is: SemiSupervised MarginBoost by F. d'AlchéBuc, Y. Grandvalet and C. Ambroise Paper. The other paper which I will briefly discuss is: EMDD: An Improved MultipleInstance Learning Technique by Qi Zhang and Sally A. Goldman Paper. 
1/23/2002
Friday 
Charles Rosenberg 
I will give an overview of a paper from the recent NIPS conference:
Partially labeled classification with Markov random walks by Martin Szummer and Tommi Jaakkola Paper. 
Last modified: Mon Apr 15 22:33:19 EDT 2002