Efficient Object Localization with Variation-Normalized Gaussianized Vectors - Intelligent Video Event Analysis and Understanding - page 107

80

75

71.2%

70

65

60

56.4%

54.8% 53.8%

54.6%
51.8%

55

47.8%

50

45

Histogram G−32 G−32N G−64 G−64N G−128 G−128N

Fig. 3 Equal Error Rates for multi-scale car detection

We can see that the Gaussianized vector representation outperforms the his-
togram of keywords in this multi-scale object detection task. In particular, using 64
Gaussian components gives the best performance. In general, normalizing against
within-class variation further improves the system.
In Figure 4, we present a few examples of correct detection and erroneous de-
tection using 64 Gaussian components. Each test image is accompanied by a “per-
feature-contribution” map. Negative and positive contributions are denoted by blue
and red, with the color saturation reflecting absolute values. The quality function
evaluated on a bounding box is the sum of all the per-feature-contributions, as dis-
cussedinSection4.
The examples of correct detection demonstrate that the system can effectively
localize one or multiple objects in complex backgrounds.
The three examples of erroneous detection probably occur for different reasons:
1) The car is a bit atypical, resulting in fewer features with highly positive contribu-
tions. 2) The two cars and some ground texture form one rectangle area with highly
positive contributions, bigger than the two true bounding boxes. 3) The car is highly
confusable with the background, resulting in too many highly negative contributions
everywhere, preventing any rectangle to yield a high value for the quality function.