Table 2 (a) Comparison with self-supervised learning and transfer learning baselines. (b) Comparison with methods using manually assisted structured labels.

(a) Comparison with self-supervised learning and transfer learning baselines

Method

0.8 k (1%)

8 k (10%)

80 k (100%)

0.1 k (1%)

1 k (10%)

10 k (100%)

All

Ours

76.6

80.9

84.6

83

88.2

90.1

83.1

Model Genesis⁵⁰

70.3

75.7

81

70.7

82.7

85.8

76

C2L⁵¹

71

76.6

82.2

75.3

83.3

85.9

77.8

Context Restoration⁹

67.8

73.9

78.7

67.9

82.4

83.8

74.6

TransVW⁷

71.2

74.3

81.7

73.6

83.8

86.2

76.1

ImageNet Pre-training⁵¹

69.8

74.4

80

69.7

82.9

84.5

74.1

p-value

8.35E-04

8.72E-04

1.94E-03

8.72E-05

4.34E-04

9.33E-04

5.88E-04

(b) Comparison with methods using human-assisted structured labels

Ours

76.6

80.8

84.8

83

88.2

90.1

82.1

LSP (Transformer)⁴¹

74.2

78.2

82.1

78.5

85.8

87.6

80.2

LSP (ConvNet)

65.8

74.5

81.9

76

85.2

87.2

80.1

DenseUN⁵⁴

75.8

72.5

71.9

72.6

82.2

77.8

80

MMGan⁵⁵

65.9

64.5

80.8

77

75.8

86.2

81.8

p-value

3.25E-03

2.89E-03

5.23E-03

3.56E-04

8.69E-04

1.05E-03

7.61E-03