IMPLEMENTATION

    The output images are as Fig R-1 . Image (a) is the current frame . Image (b) is the last frame combined with the motion arrows . Image (c) is the affine image . Using the affine parameters and the current frame to rebuild an affine image . Image (d) is the differential image which shows the difference between last image and affine image .
    In this video , the foreground is the girl in red shirt . And she’s moving to the left . The camera is also moving toward left . The motion of the background is opposite to camera . We can see that from the arrows . The major motion ( the motion of background ) is right . And the motion of the foreground object is different from background .

    Figure R-2 shows a more obvious difference between the motions of foreground and background objects . If there’s no movement from previous frame to the current frame , it’s output will be like Figure R-3 .

    This system is restricted to take a stable video as input . If the shaking of the whole video is strong , the DS algorithm can’t tell the difference from foreground and background by their own motion . See the example of an unstable video as Fig R- 4 , the arrows indicating the motions of blocks are in disorder . We can’t even tell the difference between the motion of foreground and background by observing these arrows .