Adrian Ulges

> Teaching     > Research     > Publications     > Contact    


  Current Courses (Summer 2018)

  Older Courses

  Student Theses

Concerning theses < 2013, please refer to for more information (including pdfs).

  Current Research



Content-based image search and 3D modelling are trending fields, with applications from copyright protection over mobile object search and 3D printing to community-based modeling. Key to exploiting large collections of images and 3D models, however, are efficient search and organization mechanisms. Here, SMULGRAS' goal is to investigate a multi-codal search between 2D images and 3D models: Given a camera snapshot, can we find the matching 3D model in a database? And given a 3D model, can we find images showing the corresponding object?
SMULGRAS is funded by the program Forschung für die Praxis and is executed in collaboration with Prof. Yvonne Jung from Hochschule Fulda and IntelligentGraphics GmbH.

[web demo (request access per e-mail)]  



Key to preserving (often endangered) fish species is a cheap, non-invasive sampling strategy to tackle issues such as population estimation and species association tracking. Current manual sampling methods are expensive, of limited coverage, and come with fish injury and mortality. Therefore, the goal of FIBEVID is to employ underwater computer vision technology for an automated surveying of fish. This includes the construction and deployment of suitable underwater camera rigs, as well as the development of semi-automatic localization, tracking, measuring and classification software for underwater species. FIBEVID (2016-2017) is carried out in cooperation with Ulrich Schwanecke's lab at HSRM and the National University of Sciences and Technology (NUST) in Islamabad / Pakistan. The project is sponsored by the German Academic Exchange Service (DAAD) in the program "German-Pakistan Research Cooperations".

[project website]  



Educational video is the key driver to e-learning, as it offers rich context, flexibility, and personalized learning speed. Interaction possibilities, however, are often limited to a sequential viewing. To overcome this limitation, AMIGO's goal is an interaction with videos just like with PDFs: You can search for text in e-lectures, copy text from them, navigate between pages/slides, get links to interesting stuff, etc. Key to that is an automatic indexer for e-lectures that - given a video and presentation slides - automatically localizes each slide in the video, estimating a pixel- and frame-accurate position of each text box.

[AMIGO live system (access restricted)]   [AMIGO paper]

  Other (and older) stuff

  Visual Learning from the Web

Web-based portals like Flickr or YouTube offer huge amounts of images and videos, combined with rich context (such as tags, descriptions, comments, ratings, etc). This information can be exploited for a multi-modal auto-understanding of web-based multimedia. By training visual recognition systems at larger scale and combining their output with other signals, smarter search and recommendation can be facilitated. Core research issues are learning from unreliable labels, domain drift, and new forms of supervision to learning.

[MOONVID project] [demo] [sample paper (demographics prediction)]

  Multimedia Forensics

The massive growth of multimedia data poses challenges to law enforcement, who are confronted with analysing vast amounts of evidence in digital image and video form. Scenarios are the investigation of public desasters, or of cases of child sexual abuse. Here, multimedia analysis can support forensics by content classification, similarity search, or content clustering.

[iCOP project] [sample paper (similarity search)]

  Camera-based Document Capture

Compared to scanners, 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. These challenges can be tackled by dewarping document images prior to analysis, or by robustifying optical character recognition (OCR) to take low resolution into account.

[demo] [sample paper (dewarping)]


... can be found here.

  Consultation (Sprechstunde)

Wednesdays, 17:30 - 18:30 (semester), by appointment (semester break)


Adrian Ulges
Hochschule RheinMain
Fachbereich DCSM
Studienbereich Informatik

Büro: C-201
Unter den Eichen 5
Haus C links
D-65195 Wiesbaden

e-mail: adrian dot ulges at hs-rm dot de
PGP fingerprint: E60E F8AF C1CC 6027 1C97 F35B 2E02 01DF 1FC1 CDF7