Results for DivX (MPEG-4) encoding

Below we show a comparative study on one video clip, looking at all combinations of the model parameters described above.

Saliency
Competition
Object
Segmentation
Pyramid Depth 2 Pyramid Depth 4 Pyramid Depth 6
Continuous 1 Fovea 3 Foveas 5 Foveas Continuous 1 Fovea 3 Foveas 5 Foveas Continuous 1 Fovea 3 Foveas 5 Foveas
MaxnormNone
70.7%
65.8%
70.7%
73.1%
60.1%
57.5%
61.8%
64%
56.9%
54.8%
58.3%
60.3%
FeatureMap  
76%
90%
94.3%
 
68.2%
82.6%
89.9%
 
65.4%
81.3%
89.7%
FancyOneNone
61%
65.2%
70.8%
74%
55.9%
57%
61.7%
64.4%
54.2%
54.4%
58.3%
60.8%
FeatureMap  
69.4%
78.8%
82.9%
 
60.7%
69.9%
74.2%
 
57.5%
66.6%
70.9%
FancyNone
57.3%
65.7%
71.7%
73.7%
53.6%
56.7%
62.1%
64.3%
53.1%
53.9%
58.6%
60.8%
FeatureMap  
68%
75.6%
77.8%
 
58.9%
66.2%
68.2%
 
55.7%
62.5%
64.6%

The gains in file size using our multi-foveation technique are not as high as with MPEG-1 encoding, but the trends are similar. Note that the overall quality of the original clip itself is lower with this encoder than with the MPEG-1 encoder. Pyramid depth 4 and continuous or 3 foveas seems to be an overall visually-pleasant setting.


Copyright © 2003 by the University of Southern California, iLab and Prof. Laurent Itti