OBJECT ROTATION AXIS FROM SHADING

J. Stauder, contact: J. Stauder, H. Nicolas

Goal

This demonstration is on automatic analysis of moving objects in a video sequence. The goal is the estimation of rigid object motion. More specificly, one of six degrees of freedom, the tilt angle of the object rotation axis will be estimated. The tilt angle defines the orientation of the 2D projection of the 3D rotation axis into the image plane and is a strong clue for image understanding as can be seen below.

Problem

Many known approaches on motion estimation from the literature assume diffuse illumination in the scene, such that the brightness of an object does not change by motion. But, in presence of a dominating light source, an object will be shaded. Moreover, the object shading will change if the object moves. Temporally changing shading can disturb those approaches that do not consider it. On the other side, temporally changing shading contains information on object motion and structure, than can be exploited [Pent91].

Approach

Here, object motion will be estimated from nothing than the effect of temporally changing shading, i.e. the temporal changes of brightness of the object. Input data for estimation are two images, a 2D binary object segmentation mask for the second image and a displacement vector field. No 3D information is required.

The proposed estimation method assumes the object to be rigid, to be matte, to have equally distributed surface normals and to be illuminated by a distant point light source and ambient light. For estimation, the spatial gradient of the displaced frame ratio (DFR), i.e. the frame ratio after motion compensation, is averaged inside the object silhouette. In [Stau98] is shown, that the projected object rotation axis is the mean spatial gradient of the DFR rotated by 90 degrees.

Experimental Results

The experimental results are derived as follows:

• Image sequences: Format CIF, scenes with one dominating light source
• Object segmentation: By COST 211 Analysis Model [COST]
• Displacement estimation: Block matching, half pel resolution, for each image w.r.t. the preceding image [Bier91]
• Rotation axis tilt estimation: For each image w.r.t. the preceding image

The results show that the estimated projection of the object rotation axis follows the motion of the object. Only at the beginning of the sequences, where the object segmentation mask is incomplete, the method fails.

References

[Pent91]

A.P. Pentland: Photometric motion, IEEE Trans. on Pattern Analysis and Machine Intelligence, Vol. 13, No. 9, September 1991, pp. 879-890.

[Bier91]

M. Bierling: Hierarchische Displacementschätzung zur Bewegungskompensation in digitalen Fernsehbildsequenzen, PhD thesis, University of Hannover, Germany, Fortschritt-Berichte VDI, Reihe 10, No. 179, Hannover, 1991, ISBN 3-18-147910-1.

[COST]

COST 211quat Analysis Model

[Stau98]

J. Stauder: "Object Rotation Axis from Shading", ICASSP 1999, 15.-19.3.1999, Phoenix, Arizona, USA.