Supplement to pursuit tracks chase

We use a support vector machine (SVM) to obtain a classifier that best discriminates ES and CS1. SVM requires a similarity matrix S as input. This matrix describes the pair-wise distance between all samples. In our case, the samples are the ES and CS1 targets xn. We compute similarity euclidean distance between two samples S(xm,xn) = ∑ f ∑i ∑ j(xm( f , i, j)− xn( f , i, j))2 where x( f , i, j) is the value of a pixel at the i-th position from left, j-th position from top of the window and at the f -th frame in the sequence. SVM minimizes F(w,b) = ∑m[1− ym(∑n wn ·K(xm,xn)+b)]++ γ 2 |w| 2. Function [·]+ is [x]+ = x for x > 0 and [x]+ = 0 for x < 0. K(a,b) is the Gaussian function K(a,b) = exp ( S(a,b) β ) . The variables β and γ are free parameters. β is the precision of the Gaussian function. For large β the similarity becomes K(a,b) = 1 for all a and b and the distance between two samples does not play any role. For smaller β the similarity increases more severely with constant distance increments. γ determines the complexity of the classifier (e.g. the number of support vectors). Small γ means a complex model. We use grid search and cross validation to determine β and γ that maximize the classification rate. Table 1 euclidean distance between two samples shows the results of classifier training. The table also