An Appearance-Based Prior for Hand Tracking - Advanced Concepts for Intelligent Vision Systems - page 299

Fig. 4. Time until tracking loss: comparing the original FoF to FoF with appearance cue added,
normalized to the length of every video sequence. The rightmost bars are the average over all
sequences.

Fig. 5. A wider spread of incomplete detections outperforms too narrow an influence area

spread) to test whether a larger or smaller neighborhood of higher appearance probabil-
ities would yield better results. Fig. 5 shows the tracking length on the same sequences
as above and their average.
Spreading incomplete detections only a small amount in fact hurts the performance
by, on average, 6.9% (70.48%) compared to the original FoF (75.68%, see Fig. 4).
Whereas spreading the incomplete detections to a larger area (Eq. 3) improves tracking
by 22.3% over the small spread, or 13.9% over original FoF tracking. Small-spread
suffers from poor performance early on in sequence 10. Not counting this sequence, its
performance would go up to 77.74%, which is better than original FoF tracking but still
not as good as with a wider spread.

5.2 Appearance versus Color

One might consider replacing the color cue entirely with the appearance-based proba-
bility. However, as Fig. 6 illustrates, the performance suffers significantly, whether with