• Tidak ada hasil yang ditemukan

Chapter 1 Introduction

4.5 Parameter Tuning Details

4.5.5 Min-Hash

Here we have similar parameters pertaining to the dictionary: number of words and how the dictionary is generated. In addition, we have two more parameters: T the number of hash tables andH the number of hash functions. We try 1, 5, 25, and 100 tables with 1, 2, and 3 hash functions. Based on the results in Figure 4.16, which show these combinations with different dictionary sizes and dictionary generation, we chose to go ahead with AKM dictionary with 1 M words and using 25 and 100 tables with 1 hash function. Figure 4.17 shows the results of these combinations on the four scenarios, after which we chose (T,H) = (100,1)for the benchmark. We notice the following:

• Min-Hash performance is generally worse than Inverted File. This is expected, as it is just an approximation that works best for near identical images, and not for images with large distortions [12, 13].

• Geometric consistency checks are crucial for Min-Hash, as the first step provides very poor performance.

75

102 103 104 105

30 40 50 60 70 80 90 100

bow−akmeans−10k

Number of Images

Performance (before)

102 103 104 105

10−2 10−1 100 101 102

bow−akmeans−10k

Number of Images

Search time (sec/image)

102 103 104 105

30 40 50 60 70 80 90 100

bow−akmeans−10k

Number of Images

Performance (after)

102 103 104 105

10−2 10−1 100 101 102

bow−akmeans−10k

Number of Images

Geometric Check Time (sec/image)

ivf−none−l1−l1 ivf−none−l2−l2 ivf−tfidf−l2−cos ivf−bin−l1−l1 ivf−bin−l2−l2

102 103 104 105

30 40 50 60 70 80 90 100

bow−akmeans−100k

Number of Images

Performance (before)

102 103 104 105

10−2 10−1 100 101 102

bow−akmeans−100k

Number of Images

Search time (sec/image)

102 103 104 105

30 40 50 60 70 80 90 100

bow−akmeans−100k

Number of Images

Performance (after)

102 103 104 105

10−2 10−1 100 101 102

bow−akmeans−100k

Number of Images

Geometric Check Time (sec/image)

ivf−none−l1−l1 ivf−none−l2−l2 ivf−tfidf−l2−cos ivf−bin−l1−l1 ivf−bin−l2−l2

102 103 104 105

30 40 50 60 70 80 90 100

bow−akmeans−1M

Number of Images

Performance (before)

102 103 104 105

10−2 10−1 100 101 102

bow−akmeans−1M

Number of Images

Search time (sec/image)

102 103 104 105

30 40 50 60 70 80 90 100

bow−akmeans−1M

Number of Images

Performance (after)

102 103 104 105

10−2 10−1 100 101 102

bow−akmeans−1M

Number of Images

Geometric Check Time (sec/image)

ivf−none−l1−l1 ivf−none−l2−l2 ivf−tfidf−l2−cos ivf−bin−l1−l1 ivf−bin−l2−l2

102 103 104 105

30 40 50 60 70 80 90 100

bow−hkm−1M

Number of Images

Performance (before)

102 103 104 105

10−2 10−1 100 101 102

bow−hkm−1M

Number of Images

Search time (sec/image)

102 103 104 105

30 40 50 60 70 80 90 100

bow−hkm−1M

Number of Images

Performance (after)

102 103 104 105

10−2 10−1 100 101 102

bow−hkm−1M

Number of Images

Geometric Check Time (sec/image)

ivf−none−l1−l1 ivf−none−l2−l2 ivf−tfidf−l2−cos ivf−bin−l1−l1 ivf−bin−l2−l2

Figure 4.14: Quick Tuning for Inverted File. Rows correspond to dictionaries: akm- 10K, akm-100K, 1km-1M, and hkm-1M, where the number corresponds to the number of words. First column depicts the recognition performance before geometric checks, second column shows the search time through the inverted file, third column shows the recognition performance after geometric step, while the fourth column shows the geometric check time.

Based on the results, we chose AKM dictionaries with 1 M visual words and the three combinations {tf-idf,l2,cos}, {bin,l2,l2}, and {none,l1,l1} for the full tuning in Figure 4.15.

102 103 104 105 20

40 60 80 100

Scenario 1

Number of Images

Performance (before)

102 103 104 105

10−2 10−1 100 101 102

Scenario 1

Number of Images

Search time (sec/image)

102 103 104 105

20 40 60 80 100

Scenario 1

Number of Images

Performance (after)

102 103 104 105

10−2 10−1 100 101 102

Scenario 1

Number of Images

Geometric Check Time (sec/image)

ivf−none−l1−l1 ivf−tfidf−l2−cos ivf−bin−l2−l2

102 103 104 105

20 40 60 80 100

Scenario 2

Number of Images

Performance (before)

102 103 104 105

10−2 10−1 100 101 102

Scenario 2

Number of Images

Search time (sec/image)

102 103 104 105

20 40 60 80 100

Scenario 2

Number of Images

Performance (after)

102 103 104 105

10−2 10−1 100 101 102

Scenario 2

Number of Images

Geometric Check Time (sec/image)

102 103 104 105

20 40 60 80 100

Scenario 3

Number of Images

Performance (before)

102 103 104 105

10−2 10−1 100 101 102

Scenario 3

Number of Images

Search time (sec/image)

102 103 104 105

20 40 60 80 100

Scenario 3

Number of Images

Performance (after)

102 103 104 105

10−2 10−1 100 101 102

Scenario 3

Number of Images

Geometric Check Time (sec/image)

102 103 104 105

20 40 60 80 100

Scenario 4

Number of Images

Performance (before)

102 103 104 105

10−2 10−1 100 101 102

Scenario 4

Number of Images

Search time (sec/image)

102 103 104 105

20 40 60 80 100

Scenario 4

Number of Images

Performance (after)

102 103 104 105

10−2 10−1 100 101 102

Scenario 4

Number of Images

Geometric Check Time (sec/image)

Figure 4.15: Full Tuning for Inverted File. Plots the results for the parameters chosen based on Figure 4.14.

77

102 103 104 105

0 20 40 60 80

100 bow−akmeans−10k

Number of Images

Performance (before)

102 103 104 105

10−6 10−4 10−2 100

bow−akmeans−10k

Number of Images

Search time (sec/image)

102 103 104 105

0 20 40 60 80

100 bow−akmeans−10k

Number of Images

Performance (after)

102 103 104 105

10−6 10−4 10−2 100

bow−akmeans−10k

Number of Images

Geometric Check Time (sec/image)

minhash−t1−f1 minhash−t1−f2 minhash−t1−f3 minhash−t5−f1 minhash−t5−f2 minhash−t5−f3 minhash−t25−f1 minhash−t25−f2 minhash−t25−f3 minhash−t100−f1 minhash−t100−f2 minhash−t100−f3

102 103 104 105

0 20 40 60 80

100 bow−akmeans−100k

Number of Images

Performance (before)

102 103 104 105

10−6 10−4 10−2 100

bow−akmeans−100k

Number of Images

Search time (sec/image)

102 103 104 105

0 20 40 60 80

100 bow−akmeans−100k

Number of Images

Performance (after)

102 103 104 105

10−6 10−4 10−2 100

bow−akmeans−100k

Number of Images

Geometric Check Time (sec/image)

102 103 104 105

0 20 40 60 80 100

bow−akmeans−1M

Number of Images

Performance (before)

102 103 104 105

10−6 10−4 10−2 100

bow−akmeans−1M

Number of Images

Search time (sec/image)

102 103 104 105

0 20 40 60 80 100

bow−akmeans−1M

Number of Images

Performance (after)

102 103 104 105

10−6 10−4 10−2 100

bow−akmeans−1M

Number of Images

Geometric Check Time (sec/image)

102 103 104 105

0 20 40 60 80 100

bow−hkm−1M

Number of Images

Performance (before)

102 103 104 105

10−6 10−4 10−2 100

bow−hkm−1M

Number of Images

Search time (sec/image)

102 103 104 105

0 20 40 60 80 100

bow−hkm−1M

Number of Images

Performance (after)

102 103 104 105

10−6 10−4 10−2 100

bow−hkm−1M

Number of Images

Geometric Check Time (sec/image)

Figure 4.16: Quick Tuning for Min-Hash. Rows correspond to dictionaries: akm-10K, akm-100K, akm-1M, and hkm-1M, where the number corresponds to the number of words.

First column depicts the recognition performance before geometric checks, second column shows the search time through the inverted file, third column shows the recognition per- formance after geometric step, while the fourth column shows the geometric check time.

We triedT =1, 5, 25, and 100 tables withH =1, 2, and 3 hash functions. Based on these results, we chose AKM dictionaries with 1 M visual words, T =25 and 100 tables with H=1 for full tuning in Figure 4.17.

102 103 104 105 0

20 40 60 80 100

Scenario 1

Number of Images

Performance (before)

102 103 104 105

10−2 10−1 100 101 102

Scenario 1

Number of Images

Search time (sec/image)

102 103 104 105

0 20 40 60 80 100

Scenario 1

Number of Images

Performance (after)

102 103 104 105

10−2 10−1 100 101 102

Scenario 1

Number of Images

Geometric Check Time (sec/image)

minhash−t25−f1 minhash−t100−f1

102 103 104 105

0 20 40 60 80 100

Scenario 2

Number of Images

Performance (before)

102 103 104 105

10−2 10−1 100 101 102

Scenario 2

Number of Images

Search time (sec/image)

102 103 104 105

0 20 40 60 80 100

Scenario 2

Number of Images

Performance (after)

102 103 104 105

10−2 10−1 100 101 102

Scenario 2

Number of Images

Geometric Check Time (sec/image)

102 103 104 105

0 20 40 60 80 100

Scenario 3

Number of Images

Performance (before)

102 103 104 105

10−2 10−1 100 101 102

Scenario 3

Number of Images

Search time (sec/image)

102 103 104 105

0 20 40 60 80 100

Scenario 3

Number of Images

Performance (after)

102 103 104 105

10−2 10−1 100 101 102

Scenario 3

Number of Images

Geometric Check Time (sec/image)

102 103 104 105

0 20 40 60 80 100

Scenario 4

Number of Images

Performance (before)

102 103 104 105

10−2 10−1 100 101 102

Scenario 4

Number of Images

Search time (sec/image)

102 103 104 105

0 20 40 60 80 100

Scenario 4

Number of Images

Performance (after)

102 103 104 105

10−2 10−1 100 101 102

Scenario 4

Number of Images

Geometric Check Time (sec/image)

Figure 4.17: Full Tuning for Min-Hash. Plots results for the four scenarios for the param- eters chosen in Figure 4.16. Based on these results, we choseT =100 tables withH =1 hash function.

79

Dalam dokumen Searching Large-Scale Image Collections (Halaman 90-95)

Dokumen terkait