Zisserman - 6 
Figure 2: Independent Motion Segmentation and Outlier Rejection. 
Demonstration of automatic partitioning of corner correspondences for images taken 
from a standard cinema movie, (a)-(c) Three consecutive images showing a lorry 
translating to the right as the camera pulls away. The algorithm correctly classifies 
point correspondences as (d) on the background, (e) on the lorry, and (f) rejected 
matches. 
in general this fundamental matrix will differ from the first. Consequently, objects 
undergoing different 3D motions relative to the camera may be distinguished on the 
basis of their respective fundamental matrices computed from point matches over 
two views. The problem of motion clustering is thus reduced to one of partitioning 
a set of matches into sets consistent with a fundamental matrix. As described in the 
previous section neither knowledge concerning camera calibration nor the relative 
motion between camera and object is required to estimate F. 
Computational algorithm The algorithm is a variation of that used to com 
puted the fundamental matrix for two views of a static scene. Again RANSAC 
is used, with putative fundamental matrices estimated from a random sample of 
seven points. A cluster based on this putative F is grown by determining how many 
matches over the two images are consistent with this constraint. Information about 
spatial proximity may be used to reduce the number of samplings that are neces 
sary, this is because independently moving objects usually possess spatial cohesion. 
The outcome is the set of clusters which have a greater number of correspondences 
than a threshold. Generally the background gives the largest cluster. An example 
is given in figure 2.