Global motion compensated video saliency estimation

Matthew Oakes and Charith Abhayaratne

The original video sequences and the visual saliency map video sequences estimated using our global motion compensated video visual saliency estimation methodology.

1. Tennis sequence

This short first sequence demonstrates simple saliency estimation using A = 0.75 (see paper for more information). Note that the first frame in the sequence is formed by spatial saliency estimation as there is no reference frame for comparison.

2. Hall sequence

The second sequence demonstrates how both spatial and temporal features combine to generate an overall salient representation. Spatial saliency, mainly from the hall lighting, contribute to the first few static frames after which temporal features dominant due to the local motion caused by the man entering the hall.

3. Soccer sequence

The third sequence demonstrates camera panning at different speeds. The camera motion is estimated and local object movement from the players and ball as well as small spatial features contribute to the overall visual saliency approximation.

4. Coastguard sequence

In this sequence, camera panning focuses on one of two moving boats. Note that the spatial position of the central boat remains the same but global motion compensated frame differences are exploited to estimate each vessel as locally salient in comparison to the moving background.