Symmetry plays a remarkable role in perception problems. For example, peaks of brain activity are measured in correspondence with visual patterns showing symmetries. The Discrete Symmetry Transform lets measure local symmetries in digital images. It is easy parallelizable, due to to the fact that it can be expressed as sum of convolution filters. Moreover it is fast in its pyramidal version because the level of lowest resolution at which useful information is maintained is determined and the hierarchy of symmetry is used to find interest points with different degrees of detail. Experiments on real data show the suitability of this method for real time tasks: the search of area of interest in active vision, the segmentation of complex visual patterns, the classification of sparse images and the analysis of movement. Symmetry operators have been included in JACOB (Just A COntent Based query system for video databases), a prototypal system for automatic video indexing by combining global information and local image features. The first ones derive from motion and colour analysis while the second ones (symmetry, skeleton...) depend on the class of images to be indexed. The test database containes short sequences, with images representing people and objects of various kind.

Eyes detection (image).


Face detection (image).


Multiresolution DST (image).


Keeping track of eyes (movie).


DST and optical flow (movie).

Many thanks to Professor Jitendra Malik and Dr David Beymer at Berkeley Univesity for providing the original frames of the highway.

Robotic vision (movie).