Home > Publications >

Publications

These papers are made available for personal use only, subject to author's and publisher's copyright.

This list is not complete, but it is eventually intended to contain most of my publications, some talks and posters, and some unpublished things. Another way to find articles by me is to search for my name in the LJK lab publications server. Abstracts and bibtex entries are also available for some of the older papers.

Vision and Machine Learning

  1. The BMVC'10 paper Epipolar Constraints for Multiscale Matching shows how to use Kruppa geometry to extend conventional epipolar matching to handle multi-scale interest points, giving a practically simple method that reduces false positives by a factor of 2-4.
  2. The BMVC'10 paper Multiscale Keypoint Analysis based on Complex Wavelets develops a multiscale keypoint detector based on complex wavelets.
  3. The BMVC'10 paper Feature Sets and Dimensionality Reduction for Visual Object Detection develops an improved latent SVM human detector, using an extended feature set containing HOG, LBP and LTP features and Partial Least Squares based dimensionality reduction.
  4. I have a series of papers with H. Cevikalp and others on high dimensional classification by building a simple convex model for each class, tested on visual recognition problems. All of these models are kernalizable.
    1. Our Neurocomputing Letters paper Large margin classifiers based on affine hulls shows that in some cases classifiers based on maximizing margins between affine hulls (rather than convex hulls, as for SVM) can give results as good as or better than SVM.
    2. Our CVPR'10 paper Face Recognition Based on Image Sets uses convex model constructions for face recognition from video, where each training and test example is a set of face images found by tracking.
    3. Our ICCV'08 Subspace workshop paper Large Margin Classifiers Based On Convex Class Models shows that SVM-like maximum margin classifiers separating the convex class models can also give good results.
    4. Our CVPR'08 paper Margin-Based Discriminant Dimensionality Reduction for Visual Recognition uses the below convex model constructions to heuristically find discriminant low-dimensional projections of the datasets.
    5. Our ICML'08 paper Nearest Hyperdisk Methods for High-Dimensional Classification compares models based on the bounding hyperdisk (the intersection of the affine hull and the bounding hypersphere of the training data) with affine hulls, convex hulls and bounding hyperspheres under nearest-set classification rules.
    6. Our J. Signal Proc. Systems paper Manifold Based Local Classifiers: Linear and Nonlinear Approaches studies localized and kernelized variants of the nearest affine hull and nearest convex hull classifiers, including the Local Discriminative Common Vector method - a nearest affine hull formulation that shares local variation directions between the classes.
  5. Enhanced Local Texture Feature Sets for Face Recognition under Difficult Lighting Conditions (In IEEE Trans. Image Proc. 2010). Xiaoyang Tan and B. Triggs. A study of illumination normalization, feature sets (including "Local Ternary Patterns" a generalization of LBP) and several classification strategies for aligned frontal face recognition under challenging illumination conditions. This is an extended version of our AMFG'07 (ICCV workshop) papers (paper 1, paper 2)
  6. Scene Segmentation with Conditional Random Fields Learned from Partially Labeled Images (In NIPS'07). J. Verbeek and B. Triggs. Combining Topic Models and Conditional Random Fields for semantic-level image classification. This is the CRF version of the CVPR'07 PLSA-MRF paper. It also shows how to use partial partition functions to learn a CRF from partially labelled training data.
  7. Region Classification with Markov Field Aspect Models (In CVPR'07). J. Verbeek and B. Triggs. Combining PLSA Topic Models and Markov Random Fields for semantic-level image classification. The topic model focuses the local labels on a small set that are globally meaningful for the image as a whole, and the MRF enforces local contiuity.
  8. Fast Discriminative Visual Codebooks using Randomized Clustering Forests (In NIPS'06). F. Moosmann, B. Triggs and F. Jurie. "Extremely Randomized Clustering Forest" random forest classifiers for fast patch coding in bag-of-features image classification and object localization.
  9. Sampling Strategies for Bag-of-Features Image Classification (In ECCV'06). E. Nowak, F. Jurie and B. Triggs. Experimental study of random vs. keypoint based sampling for bag-of-features image classification. The main conclusion is that although keypoint based sampling does indeed give informative patches, sample density is more important than intelligent sample positioning so denser random sampling wins out in the end.
  10. Creating Efficient Codebooks for Visual Recognition (In ICCV'05). F. Jurie and B. Triggs. Creation of efficient bag-of-words codebooks using a fixed-radius first-match-wins clusterer instead of k-means.
  11. Human Detection using Oriented Histograms of Flow and Appearance (In ECCV'06). N. Dalal, B. Triggs and C. Schmid. A variant of our CVPR'05 HOG detector that incorporates optical flow features for improved results.
  12. Hyperfeatures - Multilevel Local Coding for Visual Recognition (ECCV'06 paper, IJCV 78:15-27 2008 paper, INRIA research report RR-5655). A. Agarwal and B. Triggs. Constructing multi-level features for visual recognition by repeatedly aggregating local bag-of-feature histograms.
  13. Hierarchical Part-Based Visual Object Categorization (In CVPR'05, talk given in Sicily Object Recognition Workshop, Oct 2004). G. Bouchard and B. Triggs. Part-based object recognition using local features and hierarchies of transformations.
  14. Histograms of Oriented Gradients for Human Detection (In CVPR'05). N. Dalal and B. Triggs. An effective pedestrian detector based on evaluating histograms of oriented image gradients in a grid.
  15. On the Absolute Quadric Complex and its Application to Autocalibration (In CVPR'05). J. Ponce, T. Papadopoulo, M. Teillaud and B. Triggs. Autocalibration using line complex geometry.
  16. Monocular Human Motion Capture with a Mixture of Regressors (In IEEE Workshop on Vision for Human Computer Interaction, CVPR'05, poster, talk). A. Agarwal and B. Triggs. A multivalued regression method for recovering ambiguous 3D human pose from monocular image silhouettes.
  17. Boundary Conditions for Young - van Vliet Recursive Filtering (Correspondence to appear in IEEE Trans. Signal Processing). B. Triggs and M. Sdika. A brief technical note on avoiding boundary effects in Young - van Vliet style forwards-backwards recursions for Gaussian IIR filters.
  18. Recovering 3D Human Pose from Monocular Images (To appear in PAMI, 2005). A. Agarwal and B. Triggs. A summary of our work on sparse kernel based regression methods for recovering 3D human pose from monocular image silhouettes.
  19. Habilitation ā Diriger des Recherches - Summary of work done 2000-2004 (2.6 MB, 75pp, in French) and Collected papers 2000-2004 (12 MB, 310pp, mostly in English). Institut National Polytechnique de Grenoble, 07/01/2005. This is a kind of second doctorate that is needed to supervise doctoral students and postulate for professorships in France.
  20. Building Roadmaps of Minima and Transitions in Visual Models (To appear in IJCV in early 2005). C. Sminchisescu and B. Triggs. The journal version of our ECCV'02 roadmaps paper. Local optimization based method for finding nearby saddle points and hence nearby local minima of a convoluted high-dimensional cost surface. Applied to 3D human tracking from monocular video.
  21. Hyperdynamic Importance Sampling. C. Sminchisescu and B. Triggs. Journal version of our ECCV'02 hyperdynamics paper, to appear in J. Image & Vision Computing 2004-5, special issue for ECCV'02. MCMC sampling in a modified potential function that focuses samples on nearby saddle points. Used to find nearby local minima of a convoluted high-dimensional cost surface and applied to 3D human tracking from monocular video.
  22. Learning to Track 3D Human Motion from Silhouettes (In ICML'04). A. Agarwal and B. Triggs. A kernel regression (RVM) based approach to 3D human motion capture from monocular image silhouettes. (A tracking based extension to the below CVPR'04 paper).
  23. 3D Human Pose from Silhouettes by Relevance Vector Regression (In CVPR'04). A. Agarwal and B. Triggs. A kernel regression (RVM) based approach to recovering 3D human pose from monocular image silhouettes.
  24. Tracking Articulated Motion with Piecewise Learned Dynamical Models (In ECCV'04). A. Agarwal and B. Triggs. 2D tracking of human motion, using a scaled prismatic body model and a learned dynamical model. The dynamics is modelled by clustering training data in state space, performing a local linear dimensionality reduction for stability, training piecewise autoregressive models on the results, then iteratively reclustering and refitting to match states to their best dynamical region.
  25. Detecting Keypoints with Stable Position, Orientation and Scale under Illumination Changes (In ECCV'04). B. Triggs. A generalized form of the multi-scale Förstner-Harris keypoint detector that selects points that are maximally stable with respect to rotations, scale changes, affine deformations, and some simple types of illuminations changes, as well as being stable with respect to position changes.
  26. Kinematic Jump Processes for Monocular Human Tracking (Appeared in CVPR'03. There is also a talk given at the GDR ISIS meeting on human gestures). C. Sminchisescu and B. Triggs. A multiple hypothesis tracker for reconstruction 3D human motion from monocular video. Trapping in local minima is reduced by explicitly generating and testing the possible 3D kinematic configurations by forwards-backwards `flipping' of each body segment.
  27. Estimating Articulated Human Motion with Covariance Scaled Sampling (Special issue on Visual Analysis of Human Movement, Int. J. Robotics Research 22(6) 371-379, June 2003, 2.2 MB). C. Sminchisescu and B. Triggs. The journal version of our Covariance Scaled Sampling based human body tracker. Uses local optimization, covariance estimates and oversized sampling to build more reliable probabilistic trackers for ill-conditioned high-dimensional problems such as 3D human tracking.
  28. Building Roadmaps of Local Minima of Visual Models (Appeared in ECCV'02, poster). C. Sminchisescu and B. Triggs. Local optimization based method for finding nearby saddle points and hence nearby local minima of a convoluted high-dimensional cost surface. Applied to 3D human tracking from monocular video.
  29. Hyperdynamic Importance Sampling (Appeared in ECCV'02). C. Sminchisescu and B. Triggs. MCMC sampling in a modified potential function that focuses samples on nearby saddle points. Used to find nearby local minima of a convoluted high-dimensional cost surface and applied to 3D human tracking from monocular video.
  30. Learning to Parse Pictures of People (Appeared in ECCV'02, poster). R. Ronfard, C. Schmid and B. Triggs. Articulated 2D person detection based on learned (Support or Relevance Vector Machine) detectors for individual body parts, with dynamic programming to connect the parts into a coherent hierarchy.
  31. Camera Pose Revisited: New Linear Algorithms (In French. Appeared in RFIA'02, 256 kB). M-A. Ameller, Long Quan and and B. Triggs. Some linear methods for calibrated camera pose from 3 or 4 known 3D points.
  32. Covariance Scaled Sampling for Monocular 3D Body Tracking (Appeared in CVPR'01, 458 kB). C. Sminchisescu and B. Triggs. Using local optimization, covariance estimates and oversized sampling to build more reliable probabilistic trackers for ill-conditioned high-dimensional problems such as 3D human tracking.
  33. A Robust Multiple-Hypothesis Approach to Monocular Human Motion Tracking (INRIA Research report, 344 kB). C. Sminchisescu and B. Triggs. A summary of our initial work on estimating 3D human articular motion from monocular video, including robust feature extraction and a combined optimization + sampling framework.
  34. Joint Feature Distributions for Image Correspondence (appeared in ICCV'01. PostScript 145 kB, poster 184 kB, extended version with appendix on tensor joint image 160 kB). B. Triggs. A flexible new probabilistic approach to inter-image feature correspondence, based on explicitly modelling the joint distribution of corresponding features in several images. Here specialized to generalize the affine and projective tensorial matching constraints. Gracefully handles near-planar geometries intermediate between epipolar and plane-homograpic correspondence models. The projective case is based on a new `tensor joint image' formulation of multi-image geometry.
  35. Optimal Filters for Subpixel Interpolation and Matching (appeared in ICCV'01. PostScript 472 kB, poster 260 kB). B. Triggs. A study of linear filters for subpixel image interpolation, translation and correlation matching. The filters are optimized over a large training set of images under various error metrics and pixel spatial response functions. Emphasizes the dangers of aliasing and the critical influence of the spatial response function on accuracy.
  36. Bundle Adjustment -- A Modern Synthesis (Revised version to appear in final proceedings of Vision Algorithms'99. PostScript 1.1 MB, PDF 670 kB). B. Triggs, P. McLauchlan, R. Hartley & A. Fitzgibbon. A long survey of the various approaches to bundle adjustment, aimed at implementors in the computer vision community. By popular demand, here is a scan of Brown's 1976 survey paper The Bundle Adjustment - Progress and Prospects, D.C. Brown, Int. Archives of Photogrammetry 21(3) paper 3-03 (33 pages), 1976.
  37. Plane + Parallax, Tensors and Factorization (Final version appeared in ECCV'2000. PostScript 99 kB, PDF 136kB, slides of talk 61 kB). Studies the special form that matching tensors and their various relations take under plane + parallax, and introduces a rank 1 factorization projective SFM method based on this.
  38. Routines for Relative Pose of Two Calibrated Cameras from 5 Points (PostScript 73 kB). Technical report describing the method and performance of a multiresultant-based C library for 5 point relative orientation of two cameras.
  39. Critical Motions for Autocalibration when some Intrinsic Parameters can Vary (137 kB, PDF, revised version to appear in J. Math. Imaging & Vision, probably Vol 13:2, Oct 2000). Fredrik Kahl, Bill Triggs and Kalle Åström. Subgroup approach to autocalibration constraints, for case when some of the intrinsic parameters can vary.
  40. Camera Pose Revisited -- New Linear Algorithms (173 kB, not accepted for ECCV'00). Marc-Andre Ameller, Bill Triggs and Long Quan. New quasi-linear 4-point and nonlinear 3-point (4 solution) methods for the pose of a calibrated camera from known 3D points.
  41. A Unification of Autocalibration Methods (70 kB, appeared in ACCV 2000). Long Quan and Bill Triggs. A buggy summary of recent work on autocalibration, using the language of direction frames.
  42. Doctoral thesis (1.6 MB, 198pp, Institut National Polytechnique de Grenoble; slides 79kB). OK, so I'm a late developer. Not really a proper thesis, just a loose collection of earlier papers (in english) with a bit of introductory text (in french).
  43. Vision Algorithms'99 bundle adjustment session (My slides, 203 kB). A schematic history of bundle methods, and illustrations of typical behaviours of good and bad numerical methods for bundle problems. The bundle adjustment survey has more on this.
  44. Depth, Factorization and Plane + Parallax (69 kB). Slides of my talk at the August 1999 workshop at KTH Stockholm in honour of Jan-Olof Eklundh's 60th birthday. A summary of some of the main results of the tensor formalism for projective vision, then some new stuff on specializing it to plane + parallax (which later appeared in ECCV'00).
  45. Differential Matching Constraints (52 kB, appeared in ICCV'99, abstract). A finite difference expansion for closely spaced cameras in projective vision, used to derive differential analogues of the finite-displacement matching tensors and constraints. Much simpler than the Astrom-Heyden approach.
  46. Camera Pose and Calibration from 4 or 5 known 3D Points (54 kB, appeared in ICCV'99, abstract). Quasilinear methods for camera pose and partial calibration from 1 image of 4 or 5 known 3D points. They generalize the 6 point `Direct Linear Transform' method by incorporating some prior camera knowledge, while still allowing unknown calibration parameters to be recovered. The 4 point method recovers focal length only, the 5 point one focal length and principal point.
  47. Covariance, Gauge Freedom and all that (72 kB). Slides from my invited talk at the 1999 IEEE Workshop on Multi-View Modelling and Analysis of Visual Scenes (Fort Collins, Colorado, just after CVPR'99, organized by K. Kutulakos and A. Shashua). Discusses the way that covariance interacts with SFM's coordinate-frame gauge freedom, and how to handle this using inner constraints. Also comments on redundancy numbers and reliability. The bundle adjustment survey has much more on this.
  48. Critical Motions in Euclidean Structure from Motion (55 kB, appeared in CVPR'99, by Fredrik Kahl and myself). An investigation of the critical motions for Euclidean scene reconstruction under several common calibration constraints, using ideal theoretic algebraic geometry tools: (i) for internally calibrated orthographic and perspective cameras; (ii) in two images, for cameras with unknown focal lengths, either different or equal. Also presents numerical experiments showing the effects of near-critical configurations for the varying and fixed focal length methods.
  49. Some Notes on Factorization Methods for Projective Structure and Motion (60 kB, unpublished). A summary of the main known properties of factorization-based projective structure from motion, including the basic formulation, depth estimation and how it can sometimes be avoided, and some suggestions about statistical properties. This is not a complete paper, but just a draft of my contribution to a joint paper on SFM methods that was being written as a deliverable for the Esprit LTR project CUMULI (first author: A. Heyden), but that never got published. It might be useful to someone though.
  50. Autocalibration from Planar Scenes (522 kB, abstract, slides). Extended version of my paper in 1998 European Conference on Computer Vision, Freiburg. A direction vector based reformulation of the basic theory of autocalibration, and implementation details and experiments for conventional constant-intrinsic-parameter autocalibration from m images of a planar scene represented by inter-image homographies. m=5 for the full 5 parameter projective (f,a,s,u,v) camera model, less if some parameters are fixed a priori. In mathspeak, the basic constraint is that the projections of the two circular points of the 3D plane must lie on the image of the absolute conic. This extended version contains (among other things) 2 appendices on the SVD based planar relative orientation method used for one initialization search method, and an unused but promising homography factorization method.
  51. Optimal Estimation of Matching Constraints (98 kB, slides, extended abstract). My paper for the SMILE'98 (3D Structure from Multiple Images of Large-scale Environments) workshop after ECCV'98, to appear in Springer LNCS. Describes the general approach and initial experiments with the numerical optimization part of a modular library for matching constraint estimation. `A New Approach to Geometric Fitting' below is an earlier incarnation of part of the work.
  52. A New Approach to Geometric Fitting (89 kB, abstract). (Originally submitted to ICCV'98, Bombay). Statistical fitting of implicit parametric curves, surfaces, and algebraic relations like matching, calibration or reconstruction constraints. Explicitly finds optimal consistent estimates of the ``true underlying data points'' using efficient nonlinear constrained optimization. Allows constraints on the parameters like det(F)=0 or the Demazure constraints on the essential matrix E. Illustrated by optimal methods for F, E and the trifocal tensor. The main point is that a `direct' approach to geometric fitting based on numerical constrained optimization is simpler for the user, more general, more informative and more accurate than the currently usual `reduction' approach.
  53. Autocalibration and the Absolute Quadric (50 kB, abstract, poster, old talk). Appeared in 1997 Conference on Computer Vision and Pattern Recognition, Puerto Rico. Autocalibration and Euclidean reconstruction from an initial projective reconstruction. Based on the absolute quadric, the easy-to-use dual of the absolute conic. Nonlinear and quasi-linear methods.
  54. Linear Projective Reconstruction from Matching Tensors (42 kB, abstract, slides (41kB), revised version appeared in Image and Vision Computing). In 1996 British Machine Vision Conference, Edinburgh. Projective reconstruction methods that recover the projection matrices directly and linearly from estimated matching tensors.
  55. Projective Geometry for Image Analysis (Postscript 866kB, abstract) Roger Mohr and Bill Triggs. Tutorial on projective geometry given at International Symposium of Photogrammetry and Remote Sensing, Vienna, July 1996.
  56. Factorization Methods for Projective Structure and Motion (49 kB, abstract, poster (42kB), earlier but more detailed draft (63 kB) ). In 1996 IEEE Conf. Computer Vision and Pattern Recognition, San Francisco. Several extensions to the below factorization-based reconstruction technique, including a structure/motion factorization algorithm for lines.
  57. A Factorization Based Algorithm for Multi-Image Projective Structure and Motion (666 kB, abstract). Peter Sturm and Bill Triggs. In 1996 European Conference on Computer Vision, Cambridge, England. A practical SVD based projective reconstruction method, based on the below theory. Something like a projective version of the Tomasi-Kanade algorithm.
  58. The Geometry of Projective Reconstruction I: Matching Constraints and the Joint Image (201 kB, Abstract). Bill Triggs. (Unpublished. Submitted to Int. J. Computer Vision, Feb, 1995). Full version of (Grassmann geometry + tensors = everything) paper. The underlying geometry of multi-image projection and how it gets reflected in inter-image token matching constraints. Good stuff, but not for the index-o-phobe.
  59. Matching Constraints and the Joint Image (49 kB, abstract). Bill Triggs. In IEEE Int. Conf. Computer Vision, Cambridge, MA, June 1995. Compact summary of the above IJCV paper. Brim-full of tensors.
  60. A Fully Projective Error Model for Visual Reconstruction (66 kB, abstract). A work-in-progress that never got finished on a projective generalization of affine least squares for error modelling in computer vision. (Unpublished. June 1995. Originally submitted to ICCV'95 Workshop on Representations of Visual Scenes).

Robotics

  1. Automatic Task Planning for Robot Vision (960 kB, abstract). Bill Triggs and Christian Laugier. In 7th Int. Symposium Robotics Research, Munich, Oct 1995. Planning accessible and occlusion-free viewing positions for a robot-arm-mounted camera.
  2. Automatic Camera Placement for Robot Vision Tasks (402 kB, abstract). Bill Triggs and Christian Laugier. In IEEE Int. Conf. Robotics and Automation, Nagoya, Japan, May 1995. Earlier version of the above planner.
  3. Achieving Dextrous Grasping by Integrating Planning and Vision Based Sensing (abstract). C. Bard, C. Bellier, J. Troccaz, C. Laugier, B. Triggs and G. Vercelli. In Int. J. Robotics Research, 14, 445-64, 1995. Heuristic preshape-based grasp planning, and visual reconstruction for it.
  4. Motion Planning for Nonholonomic Vehicles: An Introduction (78 kB, some figures missing, abstract). Unpublished review lecture given at the Summer School on Computer Vision and Robotics, Newton Institute of Mathematical Sciences, Cambridge, England, June 1993.
  5. Model-Based Sonar Localization for Mobile Robots (131 kB, abstract). In Robotics and Autonomous Systems, 12 (1994), 173-186 and Int. Symp. Intelligent Robotic Systems, Zakopane, Poland, 1993. Kalman-filter based sonar localization system for mobiles, using a geometric world model. Attempts to make the most of limited sonar bandwith by building a detailed probabilistic model of each sonar event.
  6. The Oxford Robot World Model (45 kB, abstract). Bill Triggs and Stephen Cameron. In NATO ASI Expert Systems and Robotics, F-71 (1990) 275-284, Springer-Verlag. Early work on a geometric database for mobile robot world modelling. Not very conclusive.

Odds and Ends

  1. Ripley Chapter 2 and Ripley Chapter 5 slides. Transparencies for reading group on Brian Ripley's highly recommended book Pattern Recognition and Neural Networks, C.U.P. 1996. The slides cover only chapter 2, which summarizes the theory of statistical pattern recognition, and half of chapter 5 (numerical aspects of network training). With corrections incorporating some of the discussion.
  2. A Recursive Filter for Differential Observations . Unpublished summary of theory for an extension of the Kalman filter to allow differential observations (i.e. those depending on differences between states at different times rather than just state values) without requiring the traditional state-augmentation process. Designed for vehicle odometry, and will get applied when I next work on mobiles.
  3. Numerical Methods for Nonlinear Filtering . Introduction to a fragment of an unfinished summary of numerical methods for nonlinear filtering. Probably not at all useful.