Table 4 Results of multi-model training on the MBDD2025 dataset.

From: A dataset of building surface defects collected by UAVs for machine learning-based detection

Model

\({{\bf{mAP}}}_{{\bf{0.5}}}^{{\bf{test}}}\)

\({{\bf{mAP}}}_{{\bf{0.5:0.95}}}^{{\bf{test}}}\)

Crack

Leakage

Abscission

Corrosion

Bulge

\({{\bf{AP}}}_{{\bf{0.5}}}^{{\bf{test}}}\)

\({{\bf{AP}}}_{{\bf{0.5:o.95}}}^{{\bf{test}}}\)

\({{\bf{AP}}}_{{\bf{0.5}}}^{{\bf{test}}}\)

\({{\bf{AP}}}_{{\bf{0.5:0.95}}}^{{\bf{test}}}\)

\({{\bf{AP}}}_{{\bf{0.5}}}^{{\bf{test}}}\)

\({{\bf{AP}}}_{{\bf{0.5:0.95}}}^{{\bf{test}}}\)

\({{\bf{AP}}}_{{\bf{0.5}}}^{{\bf{test}}}\)

\({{\bf{AP}}}_{{\bf{0.5:0.95}}}^{{\bf{test}}}\)

\({{\bf{AP}}}_{{\bf{0.5}}}^{{\bf{test}}}\)

\({{\bf{AP}}}_{{\bf{0.5:0.95}}}^{{\bf{test}}}\)

YOLOv5-n16

86.3%

48%

76%

35.3%

95.7%

57.3%

82%

44.2%

86.9%

47.4%

90.7%

56%

YOLOv5-s16

89%

51.4%

80.2%

38%

96.4%

59.6%

84.4%

46.7%

88%

49.3%

96.1%

63.3%

YOLOv5-m16

89.3%

52%

81.1%

38.7%

96.7%

60.1%

84.9%

47.7%

87.7%

49.2%

95.9%

64.5%

YOLOv5-l16

89.4%

52.5%

80.6%

38.6%

96.8%

59.7%

85.3%

48.1%

87.1%

48.5%

97.2%

67.3%

YOLOv5-x16

89.3%

52.7%

80%

38.2%

97.2%

60.5%

85.1%

48.2%

87.2%

48.8%

97%

67.8%

YOLOv6-n17

78%

40.5%

65.4%

27.7%

93.3%

52.8%

73.3%

36.7%

81.4%

41%

76.8%

44%

YOLOv6-s17

81.8%

43.3%

71.4%

31.3%

95.1%

54.4%

76.7%

39.2%

83.8%

43.3%

82.2%

48.5%

YOLOv6-m17

82.1%

43.3%

70.1%

30.3%

94.6%

54.7%

77.3%

40%

84%

42.6%

84.2%

49.1%

YOLOv6-l17

85%

46.7%

75.9%

34.3%

95.6%

56.7%

80.5%

42.7%

85.6%

45.7%

87.5%

53.9%

YOLOv721

89.6%

50%

81%

36.6%

96.2%

56.3%

85.3%

46.5%

87.6%

46.7%

98.1%

63.8%

YOLOv7-t21

87.2%

45.4%

76.8%

32.7%

95.6%

53.7%

82.8%

42.4%

86.5%

43.8%

94.2%

54.3%

YOLOv7-x21

89.7%

50.3%

81.3%

37%

95.9%

56.4%

85%

46.5%

87.3%

47.3%

98.9%

64.5%

YOLOv8-n18

86.7%

48.6%

77.9%

36.1%

95.8%

57.7%

82.8%

44.9%

87.1%

47%%

89.8%

57.2%

YOLOv8-s18

89.2%

51.6%

81.2%

38.3%

96.8%

59.9%

84.6%

47.2%

88%

48.8%

95.3%

63.9%

YOLOv8-m18

89.4%

52.4%

81.5%

38.8%

96.8%

59.9%

84.8%

47.9%

87.9%

49.2%

95.9%

66.4%

YOLOv8-l18

89.3%

52.3%

80.3%

38.5%

96.5%

59.9%

84.5%

47.6%

87.7%

48.9%

97.3%

66.7%

YOLOv8-x18

89.4%

52.6%

81%

38.7%

96.8%

60.6%

85.3%

48.1%

87.7%

48.4%

96.4%

67.2%

YOLOv9-t22

86.5%

48.9%

77.7%

35.8%

96.2%

58.5%

81.8%

44.9%

86.9%

47.8%

90.2%

57.6%

YOLOv9-s22

89%

51.7%

80.1%

38.3%

96.6%

60.1%

84.4%

47.3%

88%

49%

95.7%

63.8%

YOLOv9-m22

89.7%

52.9%

81.2%

39.2%

96.9%

60.9%

85.1%

48.1%

88.4%

49.5%

96.8%

66.9%

YOLOv9-c22

89.4%

52.6%

80.4%

38.5%

96.8%

60.4%

85.4%

47.9%

87.1%

48.6%

97.5%

67.6%

YOLOv9-e22

89.1%

52.1%

80.3%

37.8%

96.7%

60%

84.4%

47.7%

87.2%

48.1%

96.7%

66.8%

YOLOv10-n20

86.1%

48.3%

76.5%

35.4%

95.1%

57.4%

81.1%

43.7%

84.8%

46.2%

92.9%

59%

YOLOv10-s20

88.5%

51.7%

79.7%

38.1%

96.1%

59.4%

83.8%

47.1%

87.4%

48.1%

95.3%

65.7%

YOLOv10-m20

89%

52.6%

80.6%

38.8%

96.7%

60.7%

84.9%

47.9%

86.2%

48.3%

96.6%

67.3%

YOLOv10-l20

88.1%

50%

78.5%

36.6%

95.8%

58.3%

83.3%

46.6%

87.6%

48.1%

95.1%

60.5%

YOLOv10-x20

89.3%

52.8%

80.1%

38.6%

96.5%

60.3%

85.3%

48.2%

87%

48.8%

97.5%

68.1%

YOLOv10-b20

89.1%

52.3%

80.4%

38.7%

96.8%

60.4%

84.7%

47.6%

87%

48.3%

96.6%

66.6%

YOLO11-n19

87.5%

49.3%

78.9%

36.3%

96.2%

58.5%

83.4%

45.5%

86.3%

47.6%

92.9%

58.7%

YOLO11-s19

89.3%

52.2%

81.7%

38.5%

96.8%

60.4%

84.6%

47.5%

87%

49.1%

96.4%

65.4%

YOLO11-m19

89.3%

52.8%

81.1%

38.9%

97.2%

60.8%

85.2%

48.1%

87%

49.1%

96.1%

67%

YOLO11-l19

89.4%

52.9%

80.8%

39%

97.1%

61.1%

85.3%

48.2%

87.2%

49%

96.4%

67.1%

YOLO11-x19

89.7%

53.3%

81.1%

39.3%

97%

60.8%

85.6%

48.8%

87.7%

49.5%

97.1%

68.1%

Faster R-CNN*30

83.9%

42.4%

73.6%

29.9%

92.7%

50.1%

82.2%

42.3%

83.6%

40.8%

87.6%

49%

Faster R-CNN#29

82.9%

41.1%

71.6%

29%

92.2%

49.4%

80.7%

41.1%

83.7%

41.1%

86.2%

44.7%

  1. Faster R-CNN* refers to the Faster R-CNN model with a ResNet-50 backbone, while Faster R-CNN# denotes the version using a ResNet-101 backbone.