Euclidean versus EMD error function

We tried to find out the relation between number of bins and error function. We observed an interesting behaviour (figure 4.6 on page ). For Euclidean distance error function we had the best classification rate $90\%$ for 9 bins per box. For EMD the peak was for 30 bins, the classification rate was $85.5\%$. The lack of discrimination with more bins is being solved via EMD. On the other hand with less bins these are discriminative by itself.

Table 4.3: Euclidean Distance error function and # of bins in SIFT(see figure 4.6 on page ) and how this affects the classification rate

# of bins	Euclidean Distance (success [%])	EMD (success [%])
5	85.42	72.93
7	87.50	79.17
9	89.60	81.25
11	89.60	77.08
15	85.42	79.17
20	77.08	79.17
30	75.00	85.42
45	75.00	75.00

Figure 4.6: EMD vs. Euclidean Distance Function