Game Development Reference
In-Depth Information
Table 8.4
Memory cost for each pixel (byte)
Item
RA
GMM-1
GMM-5
MS
Buffered frames
1
×
S(char)
1
×
S(char)
1
×
S(char)
M
×
S(char)
Mean values
1
×
S(float)
2
×
S(double)
2
×
S(double)
1
Weight
0
1
×
S(double)
1
×
S(double)
0
Threshold/variance
0
1
×
S(double)
1
×
S(double)
0
Match points number
0
1
×
S(char)
1
×
S(char)
0
Sum of memory
5
34
34
×
5
=
170
M
=
120
Table 8.5
Modeling time on different sequences (second)
CIF
Crossroad
Overbridge
Snowgate
Snowroad
Average
RA
1.6
1.7
1.7
1.7
1.7
GMM-1
5.9
5.9
5.8
5.8
5.9
GMM-5
11.5
11.3
10.4
10.3
10.8
MS
61.8
61.2
57.0
56.4
59.1
SD
Crossroad
Overbridge
Office
Bank
Average
RA
6.6
6.6
6.7
6.6
6.6
GMM-1
24.4
24.5
24.5
24.4
24.5
GMM-5
43.3
41.8
46.4
41.0
43.1
MS
242.8
236.2
252.8
235.9
241.9
As for the modeling time, Table
8.5
gives the detailed information for each algo-
rithm. This result indicates MS consumes the largest modeling time and GMM-5
spares much more time than RA. Moreover, it can be concluded that RA works
without dependence on scene texture, and other methods are sensitive to sequence
content.
In a brief summary, GMM-5 contributes largest to video coding performance gain
but spares a relative large memory and time cost. In practical system, especially in
parallelism or hardware environment, such GMM-5 cannot meet the requirement for
fast modeling and low memory cost. This inspires us to propose a method which can
achieve higher performance with less memory and time cost.
8.3.3 Low Cost Background Model for Video Coding
In this section, a background modeling method with low memory cost and
computational complexity is proposed to generate G-picture. For comparison, four
typical background modeling methods are implemented and embedded into the
