IP97 versus RS91

In this test, no entry/exit was allowed.

The original cost function [9] of RS91 (2) was modified for the following reason. It was observed that RS91 performed consistently better when the second term of was (2) discarded. The explanation is that the second term prefers small displacements typical for the tests done in [9]. This preference is not used by other authors as it is not realistic in many potential applications. Our test dataset has a wide range of displacements, including the large ones for which (2) is not favorable. In the experiments, we used the modified version of (2) with the second term removed.

Figure 5 shows the plots of $E_{traj}$ and $E_{link}$against $N$ ($T$) for low speeds $v= 3$ pixels. The plots compare RS91 to IP97. At low speeds, the performance of RS91 is better. Note that the algorithm RS91 needs the initial correspondences to be given, while IP97 is self-initializing.

  

Figure 5: IP97 versus RS91 for low speeds. Plots of trajectory error (left) and link error (right). `s' is the strict, `r' the relaxed criterion.

For high speeds $v= 12$ pixels (figure 6), IP97 performs better, with the difference growing with $N$. At high densities $N> 40$, the performance of RS91 deteriorates fast. The difference between the strict and the relaxed criterion-based $E_{traj}$plots characterizes the self-correction capability of an algorithm. The self-correction capabilities of RS91 and IP97 are similar. A moderate improvement in the tracking performance is achieved when the initial correspondences are given to IP97 as well.

  

Figure 6: IP97 versus RS91 for high speeds. `i' is the strict criterion with the initial correspondences given.