Computer Vision Laboratory (CVL)

Most people are very good at recognizing 3D objects, but this is still an ongoing challenge in computer vision. This course introduces the participants to state of the art in visual object recognition, with a focus on recent algorithms with good practical utility. The course consists of seven lectures, a collection of scientific papers to read, a small project, and a written exam (or an oral exam depending on the number of participants).

• Each lecture (except the first) is accompanied by a scientific paper that should be read in advance.
• For the small project, you will be given a list of projects of ample size near the end of the lecture series. Each project consists of implementing a particular algorithm described in the course, e.g. a descriptor, a tree-search method or an evaluation scheme.

Learning Goals

After the course, you should have a good grasp of the research front in visual object recognition. You should also be able to explain algorithms for the following topics: invariant frames, feature detectors, descriptor construction, matching, tree search, and performance evaluation.

The Lectures

The lectures are 90min each, and will be held at the Department of Electrical Engineering, in the room Algoritmen, starting 13:15, on the following dates: 10/9, 17/9, 24/9, 1/10, (8/10 cancelled), 15/10, 22/10, 29/10.

If you are interested in participating, you should contact Per-Erik Forssén before Sept. 1. e-mail: perfo AT isy.liu.se, Tel: 013-285654.

1. Introduction
   Object recogniton, classification, categorisation, detection, pose estimation, facial expressions. When to use local features, and why. Applications of VOR: Databases, grasping in robotics, autofocus and others. [ PDF ]
2. Image formation
   Pin-hole and thin lens models, illumination, homographies, epipolar geometry, canonical frames. [ PDF ]
3. Automatic detection of canonical frames
   DoG, affine adaptation, MSER+LAF, MSCR, EBR. [ PDF ]
4. Descriptor construction
   SIFT, SURF, ShapeContext, GeometricBlur, ContourSIFT, learning descriptors. [ PDF ]
5. Metrics for descriptor comparison
   Ratio score, Chi ² distance, EMD, visual words, bags-of-features, learning the metric. [ PDF ]
6. Tree search and hashing algorithms
   High dimensional spaces, kD-trees, BBF, ball trees, k-means tree, geometric hashing. [ PDF ]
7. Performance evaluation
   Repeatability tests, inlier-frequency curve, precision-recall and ROC curves. [ PDF ]

The Articles

• M. Brown, D. Lowe, "Invariant Features from Interest Point Groups", BMVC 2002
• K. Mikolajczyk, et al., "A Comparison of Affine Region Detectors", Springer IJCV 2006
• D. G. Lowe, "Distinctive Features from Scale-Invariant Keypoints", Springer IJCV 2004
• G. Csurka et al. "Visual Categorization with Bags of Keypoints", Workshop on Statistical Learning for CV 2004
• J.S. Beis and D. G. Lowe, "Indexing without Invariants in 3D Object Recognition", IEEE TPAMI 1999
• J. Davis and M. Goadrich, "The Relationship Between Precision-Recall and ROC Curves", ICML 2006

If you are interested in participating, you should contact Per-Erik Forssén before Sept 1. e-mail: perfo AT isy.liu.se, Tel: 013-285654.