[Other reserach: Visual Robot Control, Image Based View Synthesis ]

Saliency Maps and Attention Selection in Scale and Spatial Coordinates: An Information Theoretic Approach

By Martin Jägersand.

A more complete description of information decomposition in both scale and spatial coordinates can be found in the hypertext article: [Jag ICCV 95] . Also availible in postscript (2MB).

At an early stage in processing images it would be useful to know in what scales relevant information occurs. We use an information theoretic measure of how much information an image contains given an observer observing at a particular resolution. This is then decomposed into succesive contrasts between a series of resolution lengths. We have made experiments showing that this measure gives clear indication of characteristic scale lengths of objects in a variety of real world images.

For example, consider this image of a toy town, "Tinytown" In this image we can se structure on the macroscopic level (bars, trees, houses etc..), but there is also structure at a much smaller scale, corresponding to the textures of the ground and trees. (You may need a good monitor to see this ;). The bimodal structure of the scale space information expansion shown below accurately tells us about these two scales present in the image.

Attention in Scale and Spatial Coordinates

The following figure shows an example of how the information measure can be used for attention direction. The two upper small images show, overlaid on the original image, the information infinitesimal thresholded in spatial coordinates at two different levels. The high threshold (upper left) is adjusted as to only select a few areas. The low threshold (upper right) is set to half the high threshold to select larger regions more likely to contain a whole object. For each region selected at both thresholds, a rectangular ROI (region of interest) is generated. Some of the ROI's are shown highlighted in the center image. For each ROI, the information infinitesemal is integrated in spatial coordinates, to give the scale expansion, just as done for the whole image in the previous example, only more efficiently utilizing the precomputed R^3 information map. The scale space information decompositions corresponding to the different ROI's is shown in the small diagrams around the center image.

The end result of the attention selection, passed on for further processing, is a few selected ROI's and their scale space information.


[Other reserach: Visual Robot Control, Image Based View Synthesis ]

Back to Martin J's home page