SIFT topology

The best classification rate approx. 89% was achieved for SIFT with topology 15x9x25. However SIFT with topology 17x13 showed the similar classification power (85-88%) with number of bins between 15-30 . We saw that every SIFT had its maximum for certain number of bins and then the classification rate was lower for higher number of bins. This peak was for SIFT with higher number of tiles around 25 bins per tile, for SIFTs with lower number of tiles it was round 15 bins per tile (figure 4.9 on page ). Another interesting point was that the classification power achieved its peak for SIFT 15x9 and for other SIFTs it was not outperformed. The classification power stagnates from certain number of tiles (figure 4.11 on page ).

Another interesting observation is that the classification rate for number of bins per tile equal 10 is very low, it is lower then for surrounding number of bins: 9 and 11. The possible explanation gives us the picture figure 4.10 on page .

Table 4.5: The SIFT topology and how it influence the classification rate (see figure 4.9 on page )

SIFT Topology and its relation to the classification rate [%]

# of bins	6x3	6x4	8x4	10x5	13x9	15x9	17x13	19x15	21x17
3	28.26	36.96	44.57	30.43	50	41.3	50	-	-
5	36.96	47.83	52.17	41.3	54.35	55.43	55.43	-	-
7	45.65	56.52	63.04	45.65	60.87	71.74	-	-	-
9	53.26	58.7	64.13	50	67.39	76.09	70.65	-	-
10	46.74	45.65	53.26	51.09	64.13	60.87	67.39	-	-
11	50	54.35	60.87	60.87	75	80.43	78.26	-	-
13	55.43	61.96	65.22	59.78	80.43	81.52	82.61	-	-
15	55.43	64.13	67.39	63.04	77.17	83.7	85.87	82.61	86.96
17	53.26	70.65	66.3	64.13	79.35	84.78	86.96	83.7	85.87
21	53.26	66.3	65.22	67.39	81.52	85.87	88.04	84.78	88.04
25	48.91	66.3	68.48	75	83.7	89.13	88.04	88.04	86.96
30	-	-	-	73	80	81.25	86.96	79.35	77.17

Figure 4.9: We can see that SIFT with topology 15x9x25 performed best

Figure 4.10: Rotation and stability: (a) sample rotated $\approx 1^{\circ }$ to the left, number of bins equals to 9, (b) sample rotated $\approx 2^{\circ }$ to the right, number of bins equals to 9, (c) sample rotated $\approx 1^{\circ }$ to the left, number of bins equals to 10, (d) sample rotated $\approx 2^{\circ }$ to the right, number of bins equals to 10

Table 4.6: Increasing of SIFT tiles increases the classification rate figure 4.4 on page )

Number of tiles its relation to the classification rate [%]

SIFT Topology	Classification Rate [%]
6x3	55.43
6x4	70.65
8x4	68.48
10x5	75
13x9	83.7
15x9	89.13
17x13	88.04
19x15	88.04
21x17	88.04

Figure 4.11: SIFTs with higher number of tiles have better classification power but from certain point the classification rate stagnates