A common artifact that is observed with figure images is that objects of interest will often be illuminated in a way that causes a . curve of background-matching  intensities  along  the  interior of their projected area.  This  intensity  gradient artifact also occurs at regional borders within an inpidual object of interest. The result of figure image segmentation when this artifact occurs is a pair of segments whose bounding boxes overlap (see Fig. 2). Since these curves are usually not parallel to an axis for their entire length,  bounding  box pairs caused by this phenomenon almost always overlap. Thus, our system merges overlapping figure segments into a single segment. This merging of figure segments can sometimes  have the additional   benefit

of overcoming the occlusion of an object of interest by  a small  obstruction  (see Fig. 3).

2.5. Segmentation optimizations

2.5. 1. Segmentation subsampling. As stated thus far, the figure/ground framework provides  a  relatively fast form of object detection. However, the detection problem’s critical real-time nature is such that we would still like to identify ways in  which  the detection performance can  be  further  improved. Such an improvement is possible through the use of subsampling. Because the segmentation algorithm is currently implemented in software, each figure pixel must be considered sequentially. The subsampling optimization  improves  performance  by  considering

(a) (b)

(a) Two intersecting figure segments before they are merged

(b) The figure segment that results from their merging

Fig. 2. Merging of figure segments.

(a) (b)

(a) An occluding object (a strip of paper)

is placed over a target (a knife).

(b) Ground, current, and figure images are maintained as before.

(c) The two ends of the knife produce bounding boxes that overlap. As a result, a single figure segment is pro- duced for the target despite the occlusion.

Fig.  3.  Overcoming  an  occlusion  through  figure  segment merging.

only  every  few  pixels  instead  of  every  inpidual   between   these   two   share   that   same   state.   As a pixel. We make the assumption that if a pixel has the   result,  this  optimization  causes  object  detection    to same figureJground  state as the previously considered   fail   when   the   objects   approach   their  minimally pixel   (in   either   dimensions),   then   all   the   pixels         acceptable  height  or width. The failure  is due to the

fact that they are being pruned  away  by  the  • Figure segmentation  is only applied  to areas  where segmentation  process  (see Section  2.4)  in situations it is expected  to provide  results (i.e. in  intrisically when   the   consideration   of   each   inpidual pixel active areas or in areas currently  near an object   of would   have   resulted   in   an   acceptable   height or