丹麦哥本哈根大学计算机系 硕士课程讲义--3D计算机视觉(多视图几何)

http://isit.u-clermont1.fr/~ab/Classes/DIKU-3DCV2/

3D Computer Vision

Adrien Bartoli  >> Computer Vision classes >> 3D Computer Vision

Master class, University of Copenhagen, DIKU (Department of Computer Science)

This course is given jointly with Prof. S. Olsen. Lectures are grouped by 2. Each group is 1:30 long. More information is given on the DIKU webpage for this course.

Course Roadmap

Chapters from Hartley and Zisserman's "Multiple View Geometry" (second edition) are indicated.
The pdf files of the lecture slides (black and white handouts for printing purposes) can be downloaded by clicking on the lecture title (only after the lectures).

Lectures 1, 2 and 3 (Aug. 25, 9am)
	Introduction. What is Computer Vision?, examples, course organization, the Structure-from-Motion (SfM) paradigm, image features: points, lines, some mathematical notation and tools	Chapter 1 Appendices 4, 5
	Modeling a camera. Pin-hole and affine camera models, intrinsic and extrinsic parameters, pose. Camera calibration is postponed to Session 2.	Chapters (2), 6
	Modeling a pair of cameras. Fundamental and essential matrices, homographies, and their affine counterparts.	Chapters 9, 13, 14
Lectures 4 and 5 (Aug. 27, 9am)
	Numerical optimization I: Linear Least Squares (LLS). Basic linear algebra, pseudo-inverse, the SVD, Linear Least Squares (affine and homogeneous), Lagrange multipliers, lagrangian. Short exercise.	A5, A6 Triggs et al.
	Numerical optimization II: Nonlinear Least Squares (NLS). Gauss-Newton (GN), Levenberg-Marquardt (LM), Sequential Quadratic Programming (SQP). Short exercise.
Lectures 6 and 7 (Aug. 27, 1pm)
	Estimating two-view relationships I: Non-robust algorithms. Feature-based, SVD, 8 point algorithm, nonlinear least squares.	Chapters 4, 11, 14, A4, A5
	Estimating two-view relationships II: Robust algorithms. Overview: M-estimators, LMedS, RANSAC. Detailled description of RANSAC. Presentation of exercise 1.	Chapters 4, 11, (A6.8)
Lectures 8 to 15 given by S. Olsen
Lectures 16 and 17 (Sep. 24, 9am)
	Camera motion from two-view relationships. From the fundamental or essential matrix to the rotation and translation. From a plane homography to the rotation and translation.	Section 9.6 Faugeras et al.
	Triangulating features from two-views. Points (linear, nonlinear, optimal) and lines (exact solution).	Chapter 12
Lectures 18 and 19 (Sep. 24, 1pm)
	Projective reconstruction. The coordinate frame ambiguity. The link between a projective and a Euclidean reconstruction. Stratification.	Chapter 10
	Estimating multiple view geometry. Feature tracking. Batch / sequential / hierarchical SfM approaches. Presentation of exercise 2.	Chapter 18 Pollefeys et al.
Lectures 20 and 21given by S. Olsen
Lectures 22, 23 and 24 (Oct. 6, 9am)
	Closure-based reconstruction methods. Matching tensors. Closure constraints. Handling missing and erroneous data.	Triggs
	Bundle Adjustment. Gauss-Newton, Levenberg-Marquardt algorithms. Efficient solution to the normal equations.	Appendix A6
	Camera self-calibration from a projective reconstruction. The absolute quadric, the plane at infinity, the absolute conic and its projection.	Chapters 8, 19
Lectures 25 and 26 (Oct. 8, 9am)
	The geometry of dynamic scenes.  On coplanar on convergent linear motions.	Bartoli
	Tracking-by-detection. The principle. Keypoint detection and recognition. Deformable model fitting.	Pilet et al.
Lectures 27 and 28 (Oct. 8, 1pm)
	Image registration and 3D reconstruction for paper surfaces. 3D reconstruction, analysis of well-posedness.	Perriollat et al.
	Monocular vision in deformable environments. Deformable image registration and 4D reconstruction.	N/A

Course Description

The course will cover essential aspects of geometry based Computer Vision. Among the topics are the use of projective and affine projections, homographies, epipolar geometry, the fundamental matrix, and 3D structure computation.  Methods for camera calibration will be touched. As the title of the textbook suggests the course will focus on geometry. Practical problems, such as robust estimation for data containing blunders, are covered too. Some direct applications of these techniques that will be taught are:

• How a 3D model is built from 2D images and how synthetic images showing a new views of the scene can be rendered.

• How an artificial object may be inserted in a movie sequence such that the mixed video looks correct.

• How to construct a panorama from multiple images.

The students are expected to have a mature and operational mathematical knowledge. Linear algebra and basic geometry are mandatory disciplines. Knowledge within statistics, scientific computing, differential geometry, and numerical optimization are an advantage. The course is organized in two weeks with intensive lecturing (Calendar week 35 and 38) and two project periods (calendar weeks 36-37 and 39-41). During the project periods we will meet weakly for supervision and mutual discussion. To pass the course two projects must be made. These may be solved in groups up to three students.  Each group is supposed to write a report. The reports will be evaluated internally after the course and a decision of passed/not-passed made. The projects will be announced during the first week.  The deadlines for the reports are September 15, and October 13.