C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Current Topics in Computer Vision and
Machine Learning
B. Leibe, A. Hermans, E. Horbert, M. Kramp, W. Mehner, U. Rafi, P. Sudowe, T. Weyand
RWTH Aachen
http://www.mmp.rwth-aachen.de 19.04.2013
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Outline
◮ Organization ◮ Schedule◮ Hints for your reports
◮ Presentation of topics
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Organization
◮ Reports ◮ English or German ◮ ≈20 (±1) pages◮ bibliography counts, titlepage and TOCdon’t count ◮ LATEXis mandatory
◮ Outline required
◮ Presentation
◮ English or German ◮ ≈45 minutes
◮ Block at the end of the semester
◮ PDF, Open-/LibreOffice, Microsoft PowerPoint ◮ Templates available on our webpage
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Schedule
◮ Introductory Meeting Fri, 19.04.2013
◮ Hand in “Declaration of Compliance”1 Fri, 03.05.2013
◮ Outline due Mon, 20.05.2013
◮ Report due Mon, 24.06.2013
◮ Slides due Mon, 29.07.2013
◮ Presentation 05.08.2013 - 08.08.2013
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Hints for your report
What you should do...◮ Read andunderstand your paper
◮ Write report in yourown words
◮ Search for additional literature
◮ Take part in the library tour
◮ Compare your paper to the work of other authors
◮ Discuss (dis)advantages/problems of the methods
◮ Correctly cite all sources and state where figures have been
taken from
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Hints for your report
...and what you should not do◮ Do notcopy or translatethe original text!
◮ Do not miss thedeadlines
◮ 3d−1 penalty points for each day you exceed the deadline ◮ 50 points and youfail
◮ We will check if you ...
◮ have copied text from a paper ◮ have copied text from a website
◮ have not correctly cited all material, etc.
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Reminder
How to cite ◮ General rule◮ For every piece of information it has to be clear if it is your
own work or someone else’s
◮ Direct quote (copy original text)
Smith et al. state that their “approach combines x and y to achieve z”[5]
◮ Avoiddirect quotes ◮ Useownwords
◮ Indirect quotes
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Reminder (cont.)
How to cite◮ Mind credible sources
◮ Journal or conference papers
Peer reviewed and reliable
◮ Wikipedia
Can be altered by anyone: not a good source
◮ Use original source
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Miscellaneous
◮ Declaration of Compliance ◮ Read◮ “Ethical Guidelines for the Authoring of Academic Work” ◮ Sign and hand in the “Declaration of Compliance” until
03.05.2013
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Topic 1 - Data-driven Visual Similarity for Cross-Domain
Image Matching
T. Weyand
◮ Local-Feature based image retrieval is not invariant enough
to match images from different domains
◮ Idea: Find “unique” features stable enough to match
between domains using Examplar SVMs
◮ Applications: Painting-based Localization, Sketch-Based
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Topic 5 - Conditional Regressional Forests for Pose
Estimation
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Topic 6 - An Efficient Branch-and-Bound Algorithm for
Optimal Human Pose Estimation
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Topic 7 - Feature mining for localised crowd counting
W. Mehner◮ Obtain person counts for a grid of image cells
◮ Learn a multi-output regression from features (whole
image) to the counts of all cells
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Topic 8 - Identifying behaviors in crowd scenes using
stability analysis for dynamical systems
W. Mehner
◮ Obtain “global” information about the crowd
◮ Based purely on optical flow
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Topic 9 - Dense Tracking and Mapping in Real-Time
E. HorbertC u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Topic 10 - Extracting Object Proposals from Stereo
Images
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Topic 11 - Discriminative Decorrelation for Clustering
and Classification
P. Sudowe
◮ HOG feature dimensions are correlated
◮ Reduce this by “Whitening”
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Topic 12 - Face Detection, Pose Estimation, and
Landmark Localization in the Wild
P. Sudowe
◮ Approach to face detection and pose estimation
◮ tree-shaped dependency between landmarks
C u rr e n t T op ic s in C om p u te r V is ion an d M ac h in e L e a
Topic Overview
1. Data-driven Visual Similarity for Cross-Domain Image Matching
2. MatchMiner: Efficient Spanning Structure Mining in Large Image Collections
3. Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials
4. Decision Tree Fields
5. Conditional Regressional Forests for Pose Estimation
6. An Efficient Branch-and-Bound Algorithm for Optimal Human Pose Estimation
7. Feature mining for localised crowd counting
8. Identifying behaviors in crowd scenes using stability analysis for dynamical systems
9. Dense Tracking and Mapping in Real-Time
10. Extracting Object Proposals from Stereo Images
11. Discriminative Decorrelation for Clustering and Classification
12. Face Detection, Pose Estimation, and Landmark Localization in the Wild