Search mode

In search mode, firstly the bottom-up saliency map is computed. Additionally, we determine a top-down saliency map that competes with the bottom-up map for saliency. The top-down map is composed of an excitation and an inhibition map. The excitation map $E$ is the weighted sum of all feature maps that are important for the learned object, namely the features with weights greater than 1. The inhibition map $I$ contains the feature maps that are not present in the learned object, namely the features with weights smaller than 1:

$$\begin{array}{lcll} E & = & \frac{\sum_i(w_i * \mbox{Map}_i... ... )}{{\sum_j(w_j)}} & \quad \forall i: w_i < 1. \\ \end{array}$$

The top-down saliency map $S_{(td)}$ is obtained by: $S_{(td)} = E - I$. The final saliency map $S$ is composed as a combination of bottom-up and top-down influences. When fusing the maps, it is possible to determine the degree to which each map contributes by weighting the maps with a top-down factor $t \in [0..1]$: $S = (1-t) * S_{(bu)} + t* S_{(td)}$.

With $t = 1$, VOCUS looks only for the specified target. With $t < 1$, also bottom-up cues have an influence and may divert the focus of attention. This is also an important mechanism in human visual attention. E.g., a person suddenly entering a room catches immediately our attention, independently of the task. For the application discussed in this paper, we always use $t = 1$ and use the bottom-up saliency only to learn the weights of the training objects. Thus, the robot focuses its attention completely on the ball and not to play foul on other robots.