Nachricht

Am Dienstag, den 30.03.2010, hält Herr Jan Tünnermann um 16:30 im Raum P 1.4.17 einen Vortrag über seine Masterarbeit mit dem Titel:

Biologically inspired spatiotemporal saliency processing to enhance a computational attention model

Zusammenfassung:

In studies of attention, the bottom-up conspicuity of a visual feature is known as saliency, describing the level of difference of a feature compared to it's spatial neighbors regarding a certain dimension. Common dimensions are color, orientation or size. Traditional, computational models generate saliency maps for these dimensions and combine them into one master map of saliency from which spatial positions with high bottom-up conspicuity can be extracted.

Watanabe and Shimojo have shown that in auditory attention saliency exists with temporal aspects: In their setup a single sound can help solve a visual ambiguity while it loses this ability when embedded in an temporal sequence of similar sounds, losing it's saliency.

In modeling attention, temporal dynamics appear in mechanisms like inhibition of return, which keeps the system from analyzing the same part of the scene again. Mechanisms like that can bee seen as top-down interaction within the model.

This work proposes that temporal aspects could already start at the bottom of the system, at the point of saliency processing. Temporal saliency in vision could work similar than the previously described temporal auditory saliency.

It is observable in biological vision, that temporal events like the onset of a stimulus attract attention. Also, it is plausible, that the same onset loses this ability when embedded in a temporal sequence of similar onsets.

A first step is to investigate the existence and form of temporal saliency in biological vision. Then a traditional, computational attention model is extended to model temporal aspects of saliency with biologically plausible characteristics.

Computational models could benefit from temporal saliency by gaining information about onset signals and stimuli in motion at an early stage without forwarding to higher levels like object recognition and motion tracking.