Combining the classifier and the attention algorithm

The output of the combination of the two algorithms is the intersection of both result sets. The balls detected must be found both by the ball classifier as well as the attention algorithm. First, the foci are found in the image. Then, the classifier tries to detect balls at these specific regions. The results of the combination are shown in Table II. The test data is composed of a set of 60 realistic RoboCup images for each ball, where there is exactly one ball in each image. These were taken with backgrounds of different lighting (color) and complexity. The classifier searches areas of the first 5 foci found by the attention algorithm.

Table II: Detection rate of combined algorithm. Column 2 (attention) shows which of the 5 foci points to the ball (average).

	Att.	Classifier only		Att. and Class.
		Found	False Pos.	Found	False Pos.
red Ball	1.0	45/60	52	45/60	3
white ball	1.0	44/60	45	41/60	0
yel/red b.	1.2	57/60	63	55/60	20
Total	1.07	146/180	160	141/180	23

Figure 7: Top: Input images including round objects. Middle: False alarms in filtered images. Bottom left: False positives eliminated, ball not found. Bottom right: False detections eliminated.

The combination is very useful in eliminating false positives in images. This is shown in Fig. 7, where the false positives we were suffering from with the classifier alone are eliminated. The focus of attention is calculated in ca 1.5 sec. and the classification at these region of interest needs 200 ms (image size: 240 $\times$ 320, Pentium-M 1.7 GHz). The bulk of the running time of VOCUS is taken up by the feature computations. These may be parallelized by splitting up the processing to several CPUs [9] or with dedicated hardware [18] what makes the system real-time capable. We consider this for future work.