Dr. Adrian Ulges


Contact	Short CV	Research	Demos	Teaching and Talks	Patents	Publications	Activities	Honors	Other Links

I am a senior researcher with the Multimedia Analysis and Data Mining (MADM) Group at the German Research Center for Artificial Intelligence (DFKI) GmbH based in Kaiserslautern/Germany. My major research interests are in image and video processing, pattern recognition, multimedia forensics, and machine learning.

Contact

Dr. Adrian Ulges
Multimedia Analysis and Data Mining (MADM) Group
DFKI GmbH
Trippstadter Str. 122
D-67663 Kaiserslautern/Germany
phone: ++49 (631) 20575-4190
e-mail: adrian.ulges at dfki dot de

Short CV

I received a diploma degree in computer science (with honors) in 2005, and a PhD degree in computer science in 2009, both from the University of Kaiserslautern, Germany. Since then, I have been working with the German Research Center for Artificical Intelligence (DFKI) GmbH, where I am currently a senior researcher with the Multimedia Analysis and Data Mining (MADM) Group. I have also worked with Google Inc in 2005 (internship in Mountain View, California) and 2011 (guest scientist in Zurich, Switzerland).
My research interests are in image and video analysis, multimedia forensics, and machine learning. I have developed visual recognition systems that learn autonomously from web-based image and video content (project MOONVID) and image and video analysis technology for fighting child sexual abuse (projects iCOP, FIVES, and INBEKI). I have published over 30 scientific papers and am also active as a reviewer and committee member for several international journals, conferences, and workshops. Finally, I also have a strong background in teaching, including the lecture Multimedia Information Retrieval, 13 student theses (bachelor's, master's, diploma), and several guest lectures, lab courses, tutorials, and seminars.

Research

Visual Learning from the Web

Web-based portals like YouTube and Flickr offer huge amounts of images and videos. In the project MOONVID, we develop visual recognition approaches that autonomously learn from such web content, constructing visual models for semantic concepts, like objects ("Eiffel Tower"), activities ("soccer"), or locations ("desert"). These concepts can then be automatically detected in image and video databases, which opens up exciting applications such as semantic search, recommendation, content filtering, and a content-based targeting of ads. Our research has been awarded with a Google Research Award and a McKinsey Business Technology Award for Damian Borth.
[MOONVID project website] [demo]

Multimedia Analysis for Fighting Child Porn

Parallel to the rapid overall growth of multimedia collections, the spread of child ponography increases at alarming rates. To support police forces with the detection of child pornographic images and videos in (frequently several hundred thousands of) questioned items, we develop content analysis technology in the projects iCOP, FIVES, and INBEKI. Our approaches includes automatic porn detection and child porn detection as well as similarity search, object recognition, and intelligent video summarization.
[iCOP project website] [demo]

Motion Segmentation and Object Recognition

Another interest of mine is recognition that makes use of motion aspects in video data. For example, motion allows to segment moving objects from their background, based on which incorrect correspondences due to clutter can be discarded and the robustness and generalization of recognition methods can be improved significantly. This work has also been conducted in the project MOONVID.
[MOONVID project website]

Camera-based Document Capture

Digital cameras offer a cheap, quick, or even ubiquitious alternative for capturing documents. However, information extraction from the resulting images is hampered by low resolution, noise, and distorsion. For curled pages of books, we perform a dewarping to a flat, upright representation, using a (pseudo-)depth model of the document surface. This representation is fed to a special optical character recognition (OCR) system for noisy, warped, and low-resolution text called DIVER. In combination, this technology allows you to "google" your camera-captured document snapshots. This work was done as part of the project IPeT.
[DIVER] [dewarping demo]

Demos

Some web demos we created. For a full list of demos from our lab, please refer to madm.dfki.de.

TubeTagger is a concept-based video retrieval system that automatically learns from YouTube. The demo allows you to search a video collection using keywords, whereas all videos have been indexed completely automatically (no manual tagging was done). Smart Video Buddy is a video recommender based on computer vision technology. Semantic concepts (like "soccer game") are detected as the user is watching, and this information is used to "make videos smarter", i.e. to enrich them with adapted news, advertisements, or interesting links. Smart Video Buddy was presented at CeBIT 2010 and is the Silver Award winner at the EuroITV'10 Competition Grand Challenge. TubeFiler is an automatic genre categorizer for YouTube videos. Given a tagged YouTube clip, the system automatically sorts it into a hierarchy of genres (such as "travel->skiing"). These categories are refined further based on a visual clustering. We participated with TubeFiler in the ACM Multimedia Grand Challenge 2009.

Navidgator allows a structural browsing of image/video databases based on visual appearance. You can put certain images in your current focus of interest, zoom into the image collection, or pan over it just like a map. The demo can be run for user photographs or video frames.

This page dewarping demo illustrates how camera-captured document images can be prepared for optical character recognition. For curved book surfaces, a dewarping is applied which generates an upright, straightened representation of the page.

Teaching and Talks

Lectures

Lecture Multimedia Information Retrieval, winter term 2011/12.

Lecture Multimedia Information Retrieval, summer term 2010.

Guest Lectures and Other Talks

Tech talk Visual Learning from Flickr and YouTube: Weak Labels and User-generated Contexts at Google Mountain View, Jun 2011.
Guest lecture Content-based Image Retrieval at Thomas Breuel's course Image Processing and Image Understanding, University of Kaiserslautern, summer term 2010.
Guest Talk Visuelles Lernen von Flickr und YouTube: Nutzergenerierte Labels und Kontexte, University of Marburg, Oct 2011 (German).
Guest lecture Concept-based Video Retrieval at Dirk Krechel's course Wissens- und Contentmanagement, Hochschule RheinMain, winter term 2009/10.
Tech talk Visual Concept Learning from YouTube at Google Zurich, Dec 2009.
Guest lecture Video Processing and Compression at Daniel Keysers' course Image and Video Processing, University of Kaiserslautern, summer term 2007.
Guest lecture An Introduction to R at Thomas Breuel's course Pattern Recognition and Statistical Learning, University of Kaiserslautern, summer term 2007.

Student Theses I Supervised

For more information on all theses (including pdfs), please refer to madm.dfki.de.

Combining Social and Content Signals for Personalized Concept Detection on YouTube. Dominik Henter. Bachelor's Thesis, University of Kaiserslautern, 2012 (ongoing work).
Scalable Image Retrieval with Local Features. Ahmed Khattab. Master's Thesis, German University in Cairo and German Research Center for Artifificial Intelligence (DFKI), 2011 (ongoing work).
Ad Targeting for Web Video by Automatic Video Annotation. Markus Koch. Master's Thesis, University of Kaiserslautern, 2011.
Audio Features for Automatic Video Tagging. Dalia El Badawi. Bachelor's Thesis, German University in Cairo and German Research Center for Artifificial Intelligence (DFKI), 2011.
Local Image Features for Computer Vision. Mariam Abou-ElFadl. Bachelor's Thesis, German University in Cairo and German Research Center for Artifificial Intelligence (DFKI), 2011.
Efficient Domain Adaptation for Visual Concept Detectors. Damian Borth. Master's Thesis, University of Kaiserslautern, 2010.
Scalable Clustering for Hierarchical Content-based Browsing of Large-scale Image Collections. Tim Althoff. Bachelor's Thesis, University of Kaiserslautern, 2010.
Topic Models for Content-based Video Retrieval. Jörn Wanke. Diploma Thesis, University of Kaiserslautern, 2009.
Statistical Classification of Image Content for Visual Information Filtering. Christian Jansohn, Diploma Thesis, University of Kaiserslautern, 2009.
Shape Matching for Automatic Text Reading in Natural Scenes. Marius Renn, Diploma Thesis, University of Kaiserslautern, 2008.
Statistical Detection of Non-scene Text for Content-based Video Retrieval. Markus Koch, Bachelor's Thesis, University of Kaiserslautern, 2008.
Dewarping Documents using a Stereo Vision System. Soner Ozgun Pelvan, Master's Thesis, University of Kaiserslautern, 2007.
Stereo Reconstruction for Document Image Dewarping. Khawar Parvez, Master's Thesis, University of Kaiserslautern, 2007.

Other Teaching (Seminars, Lab Courses, Tutorials)

Some more teaching I did at the University of Kaiserslautern during my undergrad and PhD time

Bachelor Seminar Artificial Intelligence, winter term 2011/12.
Lab course Multimedia Grand Challenge - Video Categorization, summer term 2009.
Seminar Pattern Recognition, winter term 2008/09.
Lab course Video Segmentation and Recognition, winter term 2008/09.
Internship Style-based Image Annotation, Manni Duan, summer term 2008.
Tutorials for lecture Pattern Recognition and Statistical Learning, summer term 2007.
Tutorials for lecture Image and Video Processing, summer term 2007.
Internship Camera-based HCI, Andres Koetsier and Jasper Laagland, winter term 2006/07.
Lab course Screen OCR, winter term 2006/07.
Tutorials for lecture Pattern Recognition and Statistical Learning, summer term 2006.
Tutorials for lecture Image and Video Processing, summer term 2006.
Teaching assistant as an undergrad for lectures Development of Software Systems (Prof. Poetzsch-Heffter) and Numerical Algorithms (Prof. Heinrich).

Patents

US Patent (PCT/US2007072578) Recognizing Text in Images (work during 2005 internship patented by Google).

Publications

For more information on all publications (including pdfs), please refer to madm.dfki.de.

PhD Thesis

Visual Concept Learning from User-tagged Web Video. Adrian Ulges. Dissertation, Department of Computer Science, University of Kaiserslautern, 2009. (pdf, also available through Verlag Dr. Hut, ISBN 978-3-86853-248-7).

2012

Linking Visual Concept Detection with Viewer Demographics. Adrian Ulges, Markus Koch, Damian Borth. ACM Int. Conf. on Multimedia Retrieval (ICMR) 2012 (best paper award) (pdf).
Automatic Image and Video Understanding for Investigations of Child Sexual Abuse. Adrian Ulges, Christian Schulze, Armin Stahl. European Academy of Forensic Science Conference (EAFS) 2012 (extended abstract) (pdf).

2011

Learning Visual Contexts for Image Annotation from Flickr Groups. A. Ulges, M. Worring, T. Breuel. IEEE Trans. on Multimedia, 2011 (pdf).
Adapting Web-based Video Concept Detectors for Different Target Domains. D. Borth, A. Ulges, T. Breuel. Book chapter in Internet Multimedia Search and Mining, Bentham Science (to appear).
Lookapp for Ads – Content-based Advertising by Visual Concept Detection. Damian Borth, Adrian Ulges. McKinsey Business Technology Award, Kitzbühel, 2011.
Automatic Concept-to-Query Mapping for Web-based Concept Detector Training. Damian Borth, Adrian Ulges, Thomas Breuel ACM Multimedia 2011.
Scene-based Image Retrieval by Transitive Matching. A. Ulges, C. Schulze. ACM Int. Conf. on Multimedia Retrieval, 2011.
Automatic Detection of Child Pornography using Color Visual Words. A. Ulges, A. Stahl. IEEE Int. Conf. on Multimedia and Expo, 2011.
Lookapp Interactive Construction of Web-based Concept Detectors. D. Borth, A. Ulges, T. Breuel. ACM Int. Conf. on Multimedia Retrieval (demo session), 2011.
Smart Video Buddy - Content-based Live Recommendation. D. Borth, C. Kofler, A. Ulges. IEEE Int. Conf. on Multimedia and Expo (demo session), 2011.
Balanced Clustering for Content-based Image Browsing. T. Althoff, A. Ulges, D. Dengel. GI-Informatiktage 2011.

2010

Learning Automatic Concept Detectors from Online Video. A. Ulges, C. Schulze, M. Koch, T. Breuel. Computer Vision and Image Understanding 2010.
Can Motion Segmentation Improve Patch-based Object Recognition? A. Ulges, T. Breuel. Int. Conf. on Pattern Recognition (ICPR) 2010.
Relevance Filtering meets Active Learning: Improving Web-based Concept Detectors. D. Borth, A. Ulges, T. Breuel. Int. Conf. on Multimedia Information Retrieval (MIR) 2010.
Topic Models for Semantics-preserving Video Compression. J. Wanke, A. Ulges, C. Lampert, T. Breuel. Int. Conf. on Multimedia Information Retrieval (MIR) 2010.
Visual Concept Learning from Weakly Labeled Web Videos. A. Ulges, D. Borth, T. Breuel. Book chapter in Video Search and Mining, Springer-Verlag, 2010.
DFKI and University of Kaiserslautern Participation at TRECVID 2010 - Semantic Indexing Task. D. Borth, A. Ulges, M. Koch, T. Breuel. TRECVID Workshop 2010.

2009

Detecting Pornographic Video Content by Combining Image Features with Motion Information C. Jansohn, A. Ulges, T. Breuel. ACM Multimedia 2009.
Style Modeling for Tagging Personal Photo Collections. M. Duan, A. Ulges, T. Breuel, X. Wu. ACM Conf. on Image and Video Retrieval (CIVR) 2009.
TubeFiler - An Automatic Web Video Categorizer. D. Borth, J. Hees, M. Koch, A. Ulges, C. Schulze, T. Breuel, R. Paredes. ACM Multimedia 2009 - ACM Grand Challenge Track.
Fast Discriminative Linear Models for Scalable Video Tagging. R. Paredes, A. Ulges, T. Breuel. Int. Conf. on Machine Learning and Applications 2009.
TubeTagger - YouTube-based Concept Detection. A. Ulges, M. Koch, D. Borth, Thomas Breuel. Int. Workshop on Internet Multimedia Mining 2009.
DFKI-IUPR participation in TRECVID'09 High-level Feature Extraction Task. D. Borth, M. Koch, A. Ulges, T. Breuel. TRECVID Workshop 2009.
Video Copy Detection providing Localized Matches. D. Borth, A. Ulges, C. Schulze, T. Breuel. GI-Informatiktage 2009.

2008

Segmentation by Combining Optical Flow with a Color Model. A. Ulges, T. Breuel. Int. Conf. On Pattern Recognition (ICPR) 2008.
Identifying Relevant Frames in Weakly Labeled Videos for Training Concept Detectors. A. Ulges, C. Schulze, T. Breuel. ACM Conf. on Image and Video Retrieval (CIVR) 2008.
A System that Learns to Tag Videos by Watching YouTube. A. Ulges, C. Schulze, D. Keysers, T. Breuel. Int. Conf. on Vision Systems (ICVS) 2008.
A Local Discriminative Model for Background Subtraction. A. Ulges, T. Breuel. Annual Symposium of the German Association for Pattern Recognition (DAGM) 2008.
Navidgator - Similarity Based Browsing for Image & Video Databases. D. Borth, C. Schulze, A. Ulges, T. Breuel. Annual German Conference on Artificial Intelligence (KI) 2008.
Learning TRECVID'08 High-level Features from YouTube. A. Ulges, M. Koch, C. Schulze, T. Breuel. TRECVID Workshop 2008.
Multiple Instance Learning on Weakly Labeled Videos. A. Ulges, C. Schulze, T. Breuel. SAMT Workshop on Cross-Media Information Analysis and Retrieval 2008.
Keyframe Extraction for Video Taggging and Summarization. D. Borth, A. Ulges, C. Schulze, T. Breuel. GI-Informatiktage 2008.

2007

Content-based Video Tagging for Online Video Portals. A. Ulges, C. Schulze, D. Keysers, T. Breuel. MUSCLE/ImageClef Workshop 2007.
Dominant Motion Estimation using Adaptive Search of Transformation Space. A. Ulges, C. Lampert, D. Keysers, T. Breuel. Annual Symposium of the German Association for Pattern Recognition (DAGM) 2007.
Motion Interpretation using Adaptive Search of Transformation Space. A. Ulges. Technical Report, University of Kaiserslautern, 2007.

2006

Spatiogram-based Shot Distances for Video Retrieval. A. Ulges, C. Lampert, D. Keysers, T. Breuel. TRECVID Workshop 2006.
Recognizing Objects in Still Images and Video Streams. A. Ulges. Technical Report, University of Kaiserslautern, 2006.

2005

Document Image Dewarping using Robust Estimation of Curled Text Lines. A. Ulges, C. Lampert, T. Breuel. Int. Conf on Document Analysis and Recognition (ICDAR) 2005.
Indexing and Recognition of Documents Captured with a Handheld Camera. A. Ulges, Diploma Thesis, University of Kaiserslautern, 2005 (awarded by computer science faculty)
Oblivious Document Capture and Real-time Retrieval. C. Lampert, T. Braun, A. Ulges, D. Keysers, T. Breuel. First Int. Workshop on Camera-based Document Analysis and Recognition (CBDAR) 2005.

2004

Document Capture using Stereo Vision. A. Ulges and C. Lampert and T. Breuel. ACM Symposium on Document Engineering 2004.
StereoBook - Document Capture using Stereo Vision. A. Ulges. Project Thesis, TU Kaiserslautern, 2004.

Activities

Committee Memberships

Int. Workshop on Multimedia in Forensics and Intelligence (MiFor'11) (organization committee)
IEEE Multimedia, Special Issue "Multimedia in Forensics and Intelligence" (guest editor with Sebastiano Battiato, Sabu Emmanuel, Marcel Worring)
Int. Worksh. on Content Protection & Forensics (CPAF'11) (program committee)
Int. Conf. of Image Analysis and Processing (ICIAP'11) (program committee)
Int. Workshop on Multimedia in Forensics, Security and Intelligence (MiFor'10) (organization committee)
ICME Workshop on Internet Multimedia Search and Mining (IMSM'09) (program committee)

Peer Reviewing

journals: IEEE Trans. Pattern Analysis Machine Intelligence, IEEE Trans. Multimedia, IEEE Trans. Image Processing, Comp. Vision and Image Understanding, Int. J. Computer Vision, IEEE Multimedia, ELCVIA El. Letters on Comp. Vis. and Img. Analysis, Elsevier Pattern Recognition Letters, Int. J. Pattern Recognition and Artificial Intelligence, MDPI Future Internet Journal
conferences/workshops: CVPR, ECCV, ICPR, DAGM, VISAPP, MVA, DAS, CIARP, ICMLA, ICME

Honors

Jul 2010: Our research on visual learning from the web has been awarded with a Google Research Award. In a Google-funded project, we will address ad targeting for web video by concept detection.
Apr 2010: Our demo Smart Video Buddy, a content-based live recommender for video streaming, is the Silver Award winner at the EuroITV'10 Competition Grand Challenge.
Jul 2005: My diploma thesis Indexing and Recognition of Documents Captured with a Handheld Camera has been awarded by the faculty of computer science at the University of Kaiserslautern.


TubeTagger is a concept-based video retrieval system that automatically learns from YouTube. The demo allows you to search a video collection using keywords, whereas all videos have been indexed completely automatically (no manual tagging was done).	Smart Video Buddy is a video recommender based on computer vision technology. Semantic concepts (like "soccer game") are detected as the user is watching, and this information is used to "make videos smarter", i.e. to enrich them with adapted news, advertisements, or interesting links. Smart Video Buddy was presented at CeBIT 2010 and is the Silver Award winner at the EuroITV'10 Competition Grand Challenge.	TubeFiler is an automatic genre categorizer for YouTube videos. Given a tagged YouTube clip, the system automatically sorts it into a hierarchy of genres (such as "travel->skiing"). These categories are refined further based on a visual clustering. We participated with TubeFiler in the ACM Multimedia Grand Challenge 2009.	Navidgator allows a structural browsing of image/video databases based on visual appearance. You can put certain images in your current focus of interest, zoom into the image collection, or pan over it just like a map. The demo can be run for user photographs or video frames.	This page dewarping demo illustrates how camera-captured document images can be prepared for optical character recognition. For curved book surfaces, a dewarping is applied which generates an upright, straightened representation of the page.