1
INTEGRATING DEPTH IMAGE DATA INTO MARKER
CONTROLLED WATERSHED RGB IMAGE SEGMENTATION
ALGORITHM
Onur Olgac, Cihan Avci
INTRODUCTION
Segmentation of messy and crowded indoor environments based solely on RGB source images
lead to an inaccurate segmentation process and defective results, where an improvement via
auxiliary data is possible. These faulty results can be corrected through depth data acquired
together with the RGB source since it will include information about object positions in 3D and
in return, the segmentation process can be influenced positively.
The idea behind this improvement is grounded on different lighting conditions and many
occluding images since both can be sorted out if data existed to point out that it is the same
object in the lighting conditions case and that they are separate objects in the occlusion case.
THE WATERSHED CONCEPT
One can think a grayscale image as a topological map to understand the watershed concept. In
the grayscale image, blobs which are separated with high intensity values are considered as water
tanks in such an environment that rainfall effect happens. In time, the rain will fill the tanks and
eventually they will merge. Thus, the blobs will be lost as they will become one. In order to
prevent such a problem, watershed ridge lines are introduced to separate the water tanks. After
successful ridge line creation, all of the water tanks, i.e. blobs will be separated from each other.