Figure 2

Interplay between depth (disparity) and Gestalt cues in evProtoDepth saliency. The disparity maps (Row 1) have two possible depths: near (dark red) and far (light orange), and the evProtoDepth saliency (Row 2) is shown from strong (red) to weak (blue). Arranging the angle features in a convex shape generates a perceptual (proto-)object that contributes to saliency in our model. Turning any of the angles in a different orientation destroys object perception. This contribution to saliency is integrated with that resulting from differences in depth. The salience of the synthetic proto-object pattern increases as it moves closer to the camera. However, even when the proto-object moves further away in the background, it produces a strong response compared to the non-object pattern in the foreground. This demonstrates the advantage of using a proto-object model instead of directly relying on raw scene depth for nearest “object” selection by the robot. The selectivity is the strongest when the proto-object is placed closer to the camera while the non-object pattern is in the background (Column 2).