Results of the classifier alone

The ball detection cascade was learned with a total of 1000 images, with complex scenes included in the training set, and tested by using three soccer balls of different colors and patterns. The process of generating the cascade of classifiers is relatively time-consuming but it only needs to be executed once, provided a good cascade is generated. Fig. 6 shows detection results on five different kinds of balls, thus the CARTs form a correct dependency of features. Since only the upper two balls (white and yellow/red ball) and the red one given in Fig. 7 were used for learning, the figure demonstrates the classifier's ability to generalize to all balls.

For each kind of ball we ran the test with 60 images, making a total of 180 test images. The results in Table I reveal how many red, white or yellow/red balls were correctly classified or not detected, as well as the number of false positives for each ball. The problems we were facing with this approach was the difficulty to differentiate between soccer balls and other spherical objects (Fig. 7).

Figure 6: Five different kind of balls are detected by the classifier.

Table I: Detection rate of the cascade of classifier depending on the used number of stages. The cascade with 10 stages was used for the experiments with the attention system.

	# stages	Correct	Not Detected	False Pos.
red ball		52/60	8/60	114
white ball	9	48/60	12/60	70
yel/red ball		57/60	3/60	108
Total		157/180	23/180	292
red ball		45/60	15/60	52
white ball	10	44/60	16/60	45
yel/red ball		57/60	3/60	63
Total		146/180	34/180	160
red ball		45/60	15/60	51
white ball	11	42/60	18/60	47
yel/red ball		56/60	4/60	65
Total		143/180	37/180	163
red ball		44/60	16/60	26
white ball	12	29/60	31/60	31
yel/red ball		37/60	23/60	23
Total		110/180	70/180	80

The detection rate of the classifier is adjustable, i.e., a lower number of stages of the cascade increases the number of detections (hits), but also the amount of false detections. By combining the classifier and the attention algorithm the false positive detection rate will be reduced.