Adrian Ulges

Concept Learning from Web Video

Web-based portals like YouTube offer an information source from which video tagging engines can automatically learn without manual intervention. We have developed a system that autonomously learns to tag videos with semantic concepts by analyzing YouTube videos.

[project website] [demo] [press] [@heise]

Detecting Pornographic and Illegal Images and Videos

As the load of images and videos on the internet and in users' personal collections grows, the blocking of pornographic (or even illegal) image content becomes a serious issue. In the European research project FIVES, we support police forces with searching confiscated material with pattern recognition methods.

[project website]

Learning from Weakly Labeled Content

Visual Learning from the Web is difficult as tag information at flickr or YouTube is coarse and unreliable. We have developed a statistical framework that achives robustness with respect to weak labels by separating truly relevant content from non-relevant one. This is shown in the picture for the concept “basketball”, where relevant content (left) is separated from non-relevant one (right).

[publication]

Motion Segmentation and Object Recognition

Another interest of mine is the interaction between motion segmentation and object recogntion. Motion segmentation can give an segmentation of moving objects from their background. We have developed a simple recognition approach that uses motion segmentation to discard background clutter and to improve the generalization capabilities of object recognition.

[contact] [publications to come...]

OCR for Camera-based Document Capture

Imagine you capture a document by pointing your camera at it (or a camera does this automatically for you). While this simplifies capture, information extraction is more difficult due to issues like low resolution and distorsion. For this purpose, we have developed an optical character recognition (OCR) system for noisy, warped, and low-resolution text. The system was integrated with a user interface such that users can “google” in their camera-captured document snapshots.

[contact]

Document Image Dewarping

Document images delivered by digital cameras come with geometric distorsion that makes OCR difficult. This problem is difficult for curled book surfaces. To make information extraction from such images possible, we have developed novel approaches of dewarping text based on a depth model of the paper surface.

[demo] [publication] [publication]