Moreover, to obtain the same performance than the clairvoyant SQKF/S, we must consider a much larger number of particles.
The worst case theoretically for the simplest implementation is supposed to obtain the same performance whether the rotated constellation can be used or not.
It is clear that for any block-size (hence, for any subcarrier allocation scheme) we obtain the same performance in terms of outage probability.
In the next simulation, we observe that we obtain the same performance in terms of the outage probability regardless of the particular allocation scheme and the block-size, see Figure 6.
In the case of Figure 6b,c, the distance between TX and RX becomes shorter to obtain the same performance (approximately 1.4 ~ 2.2 m at a 3.75-MHz optical rate and 1.5 ~ 1.8 m at a 7.5-MHz optical rate).
The results prove that the proposed system in general was able to obtain the same performance as BIQI; in fact, BIQI was evidently outperformed when considering the RMSE indicator.
Even if a complete scalability beyond eight cores were to be assumed, nearly 2200 cores of CPU would be required to obtain the same performance as that achieved by 120 GPUs.
The extreme cases correspond to a completely corrupted visual stream, where λ A = 1, λ V = 0 and the different pose normalization techniques obtain the same performance, and to a corrupted audio signal with weights λ A = 0, λ V = 1 and the lipreading performance already observed.
Another laboratory that implements the same test would verify locally that they obtain the same performance.
Furthermore, our experiments clearly show that we can obtain the same performance as competing methods with only a fraction of the data, which in turn can save the time, money and storage capacity of deeper sequencing.
