Abb. INB-Logo

Tools + Demos

Zurück / back

Saliency by intrinsic dimensionality


We here provide additional video demos for our paper

Eleonora Vig, Michael Dorr, Thomas Martinetz, and Erhardt Barth. Intrinsic dimensionality predicts the saliency of natural dynamic scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(6):1080-1091, 2012. 


Intrinsic dimensionality of video regions


(click on above image to play video)

Illustration of the intrinsic dimensionality (iD) of videos. Using this measure, one can distinguish between homogeneous (i0D, top left), edge-like (i1D, top right) and junction-like structures (i2D, bottom left). Space-time (i.e. appearing and disappearing) corners in videos are i3D (bottom right).


A natural video and its corresponding geometrical invariants


Top left: original video; top right: invariant H; bottom left: invariant S; bottom right: invariant K.

Invariant K encodes space-time corners and non-uniform motion; Non-black locations in invariant S change in at least two spatio-temporal directions (e.g. stationary corners, straight edges that change in time); invariant H responds to at least one-dimensional changes, e.g. stationary edges or uniform regions that change in time.



     author = {Eleonora Vig and Michael Dorr and Thomas Martinetz and Erhardt Barth},
     title = {Intrinsic Dimensionality Predicts the Saliency of Natural Dynamic Scenes},
     journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
     volume = {34},
     number = {6},
     pages = {1080-1091},
     year = {2012},
     publisher = {IEEE Computer Society},