Optimal SIFT topology

Here we try to detect the peak of classification rate in relation to SIFT topology and number of bins.

Figure 4.3: Optimal Number of Bins

Table 4.1: Optimal Number of Bins (see figure 4.3 on page ) and its relation to the classification rate [%]

SIFT Topology

# of bins	6x3	6x4	6x5	8x4	10x5	13x9	15x9	17x11	17x13	19x13	19x15	21x15
3	66.67	62.5	72.92	81.25	83.33	81.25	81.25	83.33	83.33	81.25	83.33	83.33
5	70.83	54.17	70.83	85.42	83.33	83.33	81.25	83.33	83.33	83.33	85.42	85.42
7	77.08	60.42	66.67	87.5	83.33	85.42	87.5	83.33	83.33	87.5	85.42	85.42
9	79.17	56.25	66.67	89.58	85.42	89.58	89.58	85.42	85.42	87.5	85.42	87.5
10	75	37.5	81.25	79.17	87.5	83.33	81.25	85.42	85.42	87.5	87.5	87.5
11	81.25	60.42	75	89.58	87.5	89.58	89.58	87.5	87.5	89.58	89.58	87.5
13	77.08	62.5	72.92	89.58	87.5	89.58	93.75	87.5	89.58	89.58	93.75	89.58
15	72.92	41.67	75	85.42	87.5	89.58	93.75	85.42	89.58	89.58	91.67	89.58
17	72.92	43.75	75	83.33	83.33	87.5	91.67	87.5	85.42	89.58	91.67	91.67
21	66.67	47.92	72.92	79.17	81.25	87.5	87.5	87.5	83.33	87.5	91.67	89.58
25	68.75	39.58	70.83	81.25	83.33	81.25	81.25	81.25	83.33	79.17	83.33	85.42

We can see that the best classification rate was achieved when the number of bins was between 11-15. The worst case was usually when number of bins was 10 (figure 4.3 on page ).

Another interesting thing is the overall weakness of classificator with number of bins 10. The surrounding bins 9 and 11 have usually better power.

Another interesting thing is that SIFTs with topology 6x3 and 6x5 are both stronger classificators than the one with topology 6x4 eventhough the classification is better with more detailed classificators (so 6x3 should be worse than 6x4)

Table 4.2: SIFT Topology and how this influences the classification rate(see figure 4.4 on page )

SIFT Topology	Classification Rate [%]
6x3	81.25
6x4	62.50
6x5	81.25
8x4	89.58
10x6	75.00
10x5	87.50
13x9	89.58
15x9	93.75
17x11	87.50
17x13	89.58
19x13	89.58
19x15	93.75
21x15	91.67

Figure 4.4: SIFT Topology and how this influences the classification rate

We found that the classification rate is getting better as the detail of SIFT topology is increasing up to some level. Then the classification rate is decreasing (figure 4.4 on page ).