Figure 1.
Variations of visual features and co-occurring contexts in natural scenes.
(A) Similar targets occur in a variety of contexts. (B) Various targets occur in similar contexts.
Figure 2.
Patches of luminance images of natural scenes and ICs.
(A) Examples of image patches in a center-surround configuration. (B) Examples of paired center and surround ICs. (C) Examples of unpaired center ICs. (D) Examples of the ICs for the center computed alone.
Figure 3.
Patches of color images of natural scenes and ICs.
(A) Examples of color image patches in a center-surround configuration. (B) Examples of paired chromatic center and surround ICs. (C) Examples of paired achromatic center and surround ICs. (D) Examples of unpaired center ICs. (E) Examples of the ICs for the center computed alone.
Figure 4.
Probability distributions of three selected unpaired ICs.
Figure 5.
Selected paired context ICs (the left three columns of each panel) and center ICs (the right column of each panel) of 11×11×4 color patches sampled from a video database. These ICs are divided into separate red/green, blue/yellow, and bright/dark channels. (A) Selected 28 red/green or blue/yellow ICs. (B) Selected 78 bright/dark ICs. (C) Examples of unpaired center ICs.
Figure 6.
Visual saliency based on the context-mediated PDs in natural scenes.
(A) An image patch with an salient feature at the center (left), the probabilities of all ICs (middle), and the PD of the IC that has the smallest probability (right). The red circle is the probability of the central feature. (B) An image patch with an non-salient feature at the center (left), and the probabilities of all ICs (middle) and the PD of the IC that has the smallest probability (right). The red circle is the probability of the central feature.
Figure 7.
Computing visual saliency in natural scenes.
Panels illustrate the steps for computing saliency at each location in any input scene. The unpaired center ICs and the context-mediated PDs are computed beforehand from a set of natural scenes. The first step is to compute the amplitudes of the unpaired ICs for the target at each location in an input scene. The second step is to compute the saliency measure based on the context-mediated PDs in natural scenes.
Figure 8.
Examples of saliency maps of natural scenes.
First column: input scenes. Second column, saliency maps produced by our model. Third column: saliency maps given by the AIM model. Fourth column: density maps of human fixation. Saliency is coded in color-scale (red–high saliency, blue–low saliency). According to the saliency maps, the traffic lights and the cars on the road in the first scene, the red detergent box in the second scene, the pen and the stapler in the third scene, the bicycle in the fourth scene, the two men in front of the building in the fifth scene, and the stop sign in the sixth scene appear salient.
Figure 9.
ROC curve of our saliency model.
Figure 10.
Histograms of saliency measures at the random locations (green) and fixated locations in static natural scenes (blue).
Table 1.
ROC metric and KL-divergence for saliency maps of static natural scenes.
Figure 11.
Saliency maps of dynamic natural scenes.
Examples of contextual frames (the 3 left columns) and target frame (the 4th column) frames in 6 video clips and saliency maps (rightmost column). The character in the first game video, the falling water drop in the second clip, the soccer player and the ball in the third clip, the moving car and the walking policeman in the fourth clip, and the jogger in the fifth clip and the football player in the sixth clip appear salient.
Figure 12.
Histograms of saliency measures at the random locations (green) and fixated locations in dynamic natural scenes (blue).
Table 2.
KL-divergence for saliency maps of dynamic natural scenes.