Database Reference
In-Depth Information
100
90
80
70
60
50
40
30
20
Shot-to-Story
Group-to-Story
10
0
0 0 0 0 0
Average Recall (in percentage)
50
60
70
80
90
100
Fig. 7.6
Precision and recall rates of retrieval of the video stories, employing the two querying
methods: Shot-to-Story (STS) and Group-to-Story (GTS)
a 3-D graph [
193
-
195
], and has been implemented with different topologies, such
as a hierarchical graph topology [
196
] and a pyramidal hierarchical graph [
197
].
In addition, Graph Cut has also been conducted with different video descriptors,
including MPEG4 descriptors [
198
], motion vectors [
199
], and variable nonlinear
shape priors [
200
].
In order to improve performance, there are different techniques to improve Graph
Cut for video; one is to reduce the number of nodes and pixels directly, via clustering
of pixels or regions. Another technique is to firstly use a scalable or hierarchical
method for computing a simplified solution, and then iterate over to obtain more
accurate results. Tradeoffs in processing time and memory become the key issue for
long videos with a high resolution or objects with complicated object borders.
This section addresses the problem of segmentation of objects with weak edges.
Since the Graph Cut algorithm depends on the pair wise luminance of pixels, it is
dependent only on intensity distribution, and it does not take into account any shape
information of the object that should be segmented. To address this problem, an
accurate method, HOG, is introduced as a way to incorporate shape information.
The HOG was demonstrated for detecting humans in images [
201
]. HOG is
locally normalized histograms of image gradient orientations in a dense grid. It uses
the idea that the appearance and shape of local objects can often be characterized
rather well by the distribution of local intensity gradients or edge directions.
Search WWH ::
Custom Search