Motion-based video segmentation was a research focus during and shortly after the PhD of Peter Ochs at the University of Freiburg.
We contributed to the first unsupervised object-level video segmentation method that works on natural videos. Based on the framework of Brox and Malik
[BM2010],
objects are distinguished by the long-term analysis of point trajectories, which we generalized from translational to higher-order motion models. Since these trajectories yield a sparse labeling of the video, we developed robust variational methods to obtain dense segmentations. A related line of work propagates a labeling of the first frame to the whole video. For more results, including example videos, see the
segmentation page of the Computer Vision Group in Freiburg.
Publications
P. Ochs, J. Malik, T. Brox
Segmentation of moving objects by long term video analysis
IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(6):1187-1200, 2014. PDF
P. Ochs, T. Brox
Higher Order Motion Models and Spectral Clustering
IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), 2012. PDF
P. Ochs, T. Brox
Object segmentation in video: a hierarchical variational approach for turning point trajectories into dense regions
IEEE International Conference on Computer Vision (ICCV), 2011. PDF
N.S. Nagaraja, P. Ochs, K. Liu, T. Brox
Hierarchy of Localized Random Forests for Video Annotation
In A. Pinz, T. Pock, H. Bischof, F. Leberl (Eds.): Pattern Recognition (Proc. DAGM). Lecture Notes in Computer Science, Vol. 7476, Springer, 2012. PDF
S. Müller, P. Ochs, J. Weickert, N. Graf
Robust interactive multi-label segmentation with an advanced edge detector
In B. Andres, B. Rosenhahn (Eds.): German Conference on Pattern Recognition (GCPR). Lecture Notes in Computer Science, Vol. 9796, 117-128, Springer, 2016. PDF
