Table 10 Performance (Recall@k) Comparison under Zero-shot setting.

From: Learning optimal image representations through noise injection for fine-grained search

Method

CARS-196

Cub-200-2011

k = 1

k = 2

k = 4

k = 8

k = 16

k = 1

k = 2

k = 4

k = 8

k = 16

Triplet7

39.1

50.4

63.3

74.5

84.1

36.1

48.6

59.3

70.0

80.2

LiftedStruct31

49.0

60.3

72.1

81.5

89.2

47.2

58.9

70.2

80.2

89.3

N-pairs32

53.9

66.8

77.7

86.3

-

45.4

58.4

69.5

79.4

-

SCDA27

58.5

69.8

79.1

86.2

91.8

62.2

74.2

83.2

90.1

94.3

CRL-WSL28

63.9

73.7

82.1

89.2

93.7

65.9

76.5

85.3

90.3

94.4

DGCRL34

75.9

83.9

89.7

94.0

96.6

67.9

79.1

86.2

91.8

94.8

EPSHN50

82.7

89.3

93.0

-

-

64.9

75.3

83.5

-

-

Zheng et al.37

81.1

88.8

93.7

96.7

-

55.2

68.7

79.0

89.5

-

Duan et al.38

78.2

86.2

92.0

95.5

-

61.2

73.7

83.3

90.3

-

Yingying et al. (VGG16-based)30

73.2

82.1

88.6

93.2

95.4

67.5

78.2

86.7

92.0

95.1

D & C39

87.76

70.67

65.97

-

-

68.16

69.49

55.35

-

-

Yingying et al. res101-based)30

85.4

91.2

94.4

96.5

97.7

73.1

81.5

86.6

92.7

95.4

McSAP52

84.6

91.5

95.1

97.4

-

63.5

75.6

84.8

91.3

-

Adaptive hierarchical53

82.4

89.5

93.8

95.9

-

65.3

76.1

84.7

90.7

-

HSE-EPSHN54

85.4

91.2

96.9

-

-

66.9

77.4

85.5

-

-

HSE-PA54

89.6

93.8

96.0

-

-

70.6

80.1

87.1

-

-

Multi-Proxy55

90.3

93.7

96.3

-

-

69.6

79.9

87.0

-

-

Anti-Collapse56

90.5

94.6

-

-

-

71.7

81.2

-

-

-

NormSoftmax51251

84.2

90.4

94.4

96.9

-

61.3

73.9

83.5

90.0

-

NormSoftmax204851

89.3

94.1

96.4

98.0

-

65.3

76.7

85.4

91.8

-

Our Method (R18)512

85.75

91.64

94.91

96.88

98.70

63.37

74.14

83.24

90.51

94.25

Our Method (R50)512

87.27

92.74

95.87

97.70

98.83

66.81

77.14

85.01

91.24

94.55

Our Method (R50)1024

88.17

93.36

96.27

97.82

99.03

67.34

77.57

85.57

91.27

95.59

Our Method (R50)2048

89.78

94.43

96.70

98.22

98.94

68.60

78.95

86.83

91.90

95.0

Our Method (R101)512

90.22

94.34

96.65

98.16

98.94

69.04

79.29

86.66

92.10

95.44

Our Method (R101)2048

91.33

95.20

97.34

98.55

99.13

71.59

81.62

88.18

92.91

95.90

  1. Significant values are in [bold].