4.2.2
This experiment compares the AM-estimator with the M- estimator in computing relative rotation of motion sequences. Consider a sequence of images in Fig.4.6. Corners can be detected from these images as in Fig.4.7 using the Wang-Brady detector [Wang and Brady 1991]. The data
where and
, is a set of matched point
pairs between two images. A previous work [Wang and Li 1994] shows
that when the rotation axis
is known, a unique solution can be computed using only one pair of
corresponding points and the LS solution can be
obtained using m pairs by minimizing
where
A unique solution exists for the LS problem [Wang and Li 1994]. It is determined by the equation
where , that is,
The above formulation is based on an assumption that all the pairs
are correct correspondences. This may not
be true in practice. For example, due to acceleration and deceleration,
turning and occlusion, the measurements can change drastically and false
matches, i.e.
outliers, can occur. The LS estimate can get arbitrarily
wrong when outliers are present in the data d. When outliers are
present, the M-estimator can produce the more reliable estimate than the
LS estimator. The AM-estimator further improve the
M-estimator to a significant extent.
The AM-estimator minimizes, instead of (4.17), the following
where is an adaptive potential function. By setting
and using
,
one obtains
Rearranging the above gives the following fixed-point equation
where . It is solved iteratively
with decreasing
value.
Figs.4.8 and 4.9 show the estimated rotation angles (in the vertical direction) between consecutive frames (the label on the horizontal axis is the frame number) and the corresponding standard deviations (in vertical bars) computed using the LS-, M- and AM-estimators. From Fig.4.8, we see that with 20% outliers, the M-estimator still works quite well while the LS estimator has broken down. In fact, the breakdown point of the LS estimator is less than 5%.
From Fig.4.9, we see that M- and AM-estimators are comparable when the data contains less than 20% of outliers. Above this percentage, the AM-estimator demonstrates its enhanced stability. The AM-estimator continues to work well when the M-estimate is broken down by outliers. The AM-estimator has the breakdown point of 60%. This illustrates that the AM-estimator has a considerably higher actual breakpoint than the M-estimator.