Automatic/semi-automatic 3-D layering

The demonstrations given below illustrate video segmentation of a scene in foreground and background areas. The foreground could be composed of several objects, and the background could be further segmented in regions of different depth.


This work was supported in part by the EC (IT 22012), and the results presented below is the ICS-FORTH contribution to the NEMESIS demonstration video tape (delivered on May 31, 1997). The ICS-FORTH team was the following:

  • G. Tziritas, Associate Professor, scientific responsible
  • N. Alvertos, Assistant Professor
  • N. Komodakis, graduate student
  • I. Grinias, research assistant
  • G. Tzanetakis and M. Traka, undergraduate students

Static camera with moving objects

The detection of the moving objects is based on the temporal change of the image luminance. Two particular cases are distinguished concerning the background: a reference to the whole background is available, or not. If this reference exists two modules are needed: change detection between the current and the reference frames, and update of the reference frame. If the reference frame is not available, the change detection is performed between the current and the previous frames.
The original sequence is the MPEG-4 Hall Monitor test sequence, with low spatial details and low amount of movement. The moving persons are given in the foreground layers sequence.
For further information please see the following publication:
N. Paragios and G. Tziritas, Detection and location of moving objects using deterministic relaxation algorithms, Intern. Conf. on Pattern Recognition, Vol. I, pp. 201-205, Austria, 1996.
Acknowledgment: M. Traka contributed preciously in conducting the experiments.

Camera moving in a static scene

The depth layering will be obtained in the behalf of the motion parallax phenomenon. The only condition is that the 3-D camera motion is translational. The algorithm used here is composed of three modules: 2-D motion estimation, test of the 3-D translation hypothesis, and motion field segmentation. The original sequence is the MPEG-2 Flower Garden test sequence. A car is inserted between two extracted depth layers in the manipulated video. For further information please see the following publication:
N. Komodakis and G. Tziritas, Robust 3-D motion estimation and depth layering, Intern. Conference on Digital Signal Processing, Santorini, 1997.
Acknowledgment: G.Tzanetakis has implemented the video manipulation experiment.

Moving camera with moving objects

A semi-automatic segmentation method is used in this case. It is composed of two modules: global 2-D parametric motion estimation, and seeded region growing segmentation. In addition the moving objects are tracked along the image sequence.


  • The original sequences are the following MPEG-4 test sequences:
  • Coast Guard sequence, with medium spatial details and low amount of movement
  • Foreman sequence, with low spatial details and medium amount of movement
  • The extracted foreground layers are given in the following sequences:
  • Coast Guard foreground layers
  • Foreman foreground layer

For further information please see the following publication:
I. Grinias and G. Tziritas, A semi-automatic seeded region growing algorithm for video object segmentation and tracking. (in preparation)