Semantic Scene Analysis

TNT members involved in this project:
Hanno Ackermann, Ph.D.
Yuren Cong, M.Sc.
Wentong Liao, M.Sc.
Prof. Dr.-Ing. Bodo Rosenhahn
Frederik Schubert, M. Sc.

Scene understanding is a challenging topic in computer vision, robots and artificial intelligence. Given one or more images, we want to infer what type of scene is shown in the image, what objects are visible, and physical or contextual relations between the observed objects. This information is important in many applications, such as robot navigation, image search, or surveillance applications.

Relations between objects can be given by physical information, such as "in front of " or "above". More abstractly, however, humans usually consider implicit relations between objects: For instance, both a table and the chairs around the table are "above" the floor. A human observer, on the other hand, would rather consider them to be a single group of objects. In other words, table and chairs define a relation which is more than just "in front of "or "next to". This type of implicitly defined additional information is what we consider as semantic or contextual information.

We estimate semantic information defined between objects in the scene, and construct a so-called scene graph. Scene graphs neatly represent all the objects within a scene, and allow to analyze the content of an image, or to even compare two images semantically, i.e. with respect to their contents and the relations between their objects.

 

Figure 1: Example of an observed scene (left) and the scene graph constructed from it (right).

If you are looking for an interesting topic for you bachelor or master thesis, please contact Wentong Liao or Hanno Ackermann.

If you are looking for a topic for your Master or Bachelor thesis, and you are interested in analyzing and modelling abstract problems, please do not hesitate to contact Wentong Liao or Hanno Ackermann. You are required to have good programming skills (MatLab, Python, Java or C++) and you need a good understanding of, for instance, linear algebra or statistics.

We provide a GUI implemented in Matlab for generating ground truth scene graphs and visualising the generated graphs.
It contains the manually labeled scene graph data of NYU_V2_dataset. For more details please refer to the readme in the file.

Show recent publications only
  • Conference Contributions
    • He Sen, Liao Wentong, Hamed Rezazadegan Tavakoli, Michael Ying Yang, Bodo Rosenhahn, Nicolas Pugeault
      Image Captioning through Image Transformer
      Asian Conference on Computer Vision (ACCV), IEEE, Kyoto, November 2020
    • Yuren Cong, Hanno Ackermann, Wentong Liao, Michael Ying Yang, Bodo Rosenhahn
      NODIS: Neural Ordinary Differential Scene Understanding
      European Conference on Computer Vision (ECCV), August 2020
    • Christoph Reinders, Hanno Ackermann, Michael Ying Yang, Bodo Rosenhahn
      Object Recognition from very few Training Examples for Enhancing Bicycle Maps
      2018 IEEE Intelligent Vehicles Symposium (IV), June 2018
    • Florian Kluger, Hanno Ackermann, Michael Ying Yang, Bodo Rosenhahn
      Deep Learning for Vanishing Point Detection Using an Inverse Gnomonic Projection
      39th German Conference on Pattern Recognition, Springer Lecture Notes in Computer Science (LNCS), Basel, Switzerland, September 2017
    • Wentong Liao, Chun Yang, Michael Ying Yang, Bodo Rosenhahn
      Security Event Recognition for Visual Surveillance
      ISPRS Annals of Photogrammetry, Remote Sensing \& Spatial Information Sciences, Vol. 4, June 2017
  • Journals
    • Michael Ying Yang, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn
      On support relations and semantic scene graphs
      ISPRS Journal of Photogrammetry and Remote Sensing, Elsevier, Vol. 131, pp. 15-25, July 2017