Information Technology Reference
In-Depth Information
Fig. 6.
Contribution of appearance versus color. Shown are the averages over all sequences.
or without a larger “spread.” One notes also that spreading incomplete detections with-
out combining with the color probability exhibits even worse performance, and only in
the combination into a three-cue versus a two-cue tracker does the appearance-based
FoF tracker's full potential get realized.
6
Discussion
Most of the performance improvement can be attributed to the appearance cue providing
useful information when the color segmentation fails and considers nearly every pixel
as skin-colored. If LK features are lost or in violation of the flocking constraints during
those cases, the appearance cue limits the placement of re-located features to likely
hand pixels, instead of landing on background artifacts that happen to be skin-colored.
The color modality generally fails if the tracking initialization is poor (no good match
between observed hand location and mask), and if extensive camera motion changed
the composition of the background color.
A wider spread considers rather more than fewer pixels of hand appearance, due to
the imprecise segmentation achieved. Hence, the color cue is often still very important,
particularly with very cluttered backgrounds in which the hand detection returns rather
high scores. Given these considerations, failures most frequently occur when the hand
posture changes in front of skin-colored, cluttered backgrounds.
Appearance on its own is currently an inferior cue to color. If a good color histogram
is learned during tracking initialization, it provides an excellent and very precise cue
for which pixels belong to the hand and which do not. As articulated object detection
improves, we expect the appearance cue to become more important. Equally, segmen-
tation during detection (e.g., [10,18]) can supply probabilistic segmentation with better
resolution than Eq. 3-5, in turn improving the value of the appearance cue.
Traditional FoF tracking favors objects with more distinct and more uniform color.
Tracking with this extenion to FoF eccels in performance if the object has a distinct
appearance.
7
Conclusions
The power of the traditional FoF tracking lies in its combination of two image cues
so that it can continue to track successfully even if only one “constancy” assumption
Search WWH ::
Custom Search