Information Technology Reference
In-Depth Information
Fig. 12
Three test images and the corresponding depth map, where the bottom right is the
depth map and the remainder refers to the original images
6
Evaluation of Different Methods
In the previous sections, we briefly summarised the principles and characteristics
of a few classical and typical stereo matching algorithms. To evaluate their perfor-
mance in information retrieval, we in this section compare these algorithms using
the RUSHES database that was designed for video retrieval [3]. This database con-
sists of 134 raw videos and has about 14 hours' length in total, provided by Span-
ish communication group Euskal Irrati Telebista (EITB). The videos used in this
database include various contents, e.g. interviews, football matches, aerial views,
shopping and rowing, etc. 3064 key frames have been extracted from the overall
videos.
To illustrate the comparisons we here use exemplar image pairs taken from the
database. Fig. 14 shows five image pairs that denote different image backgrounds.
For example, 1st row shows an aerial view of a building from a helicopter, while
row 4 reveals shots from a moving hand-held video camera. These scenes are chal-
lenging to the existing stereo correspondence algorithms in the sense that a number
