Table 5 Comparison between base method (Method 3+M+D) and ensembled variants.
From: Inspection of railway catenary systems using machine learning with domain knowledge integration
YOLO11 based methods | |||||
|---|---|---|---|---|---|
AP50 | Recall | Precision | F1 | Time | |
Basic elements | Full inference | ||||
Base method | 89.26 (1.68) | 91.45 (1.61) | 85.55 (1.61) | 88.36 (1.28) | 0.894 (0.018) |
Default TTA | 86.68 (1.64) | 89.64 (1.35) | 80.33 (1.93) | 84.64 (1.32) | 1.483 (0.112) |
Custom TTA | 88.82 (1.36) | 90.63 (1.32) | 89.67 (1.00) | 90.13 (0.77) | 1.646 (0.076) |
Ensembled (T=4) | 90.12 | 91.52 | 92.66 | 92.06 | 3.195 |
Small elements | |||||
Base method | 69.24 (1.88) | 88.94 (1.35) | 56.66 (1.78) | 69.17 (1.48) | |
Default TTA | 69.58 (2.33) | 89.55 (0.68) | 50.40 (2.21) | 64.30 (1.92) | |
Custom TTA | 68.88 (1.43) | 86.75 (1.37) | 60.10 (3.22) | 70.89 (2.22) | |
Ensembled (T=4) | 70.30 | 88.36 | 61.55 | 72.53 | |
RT-DETR based methods | |||||
|---|---|---|---|---|---|
AP50 | Recall | Precision | F1 | Time | |
Basic elements | Full inference | ||||
Base method | 90.73 (2.16) | 93.60 (1.59) | 72.26 (3.47) | 81.41 (2.56) | 1.731 (0.040) |
Default TTA | 87.65 (2.22) | 90.59 (1.74) | 64.76 (3.95) | 75.26 (3.13) | 3.144 (0.063) |
Custom TTA | 90.57 (1.12) | 92.85 (0.84) | 77.38 (3.04) | 84.26 (1.97) | 2.946 (0.070) |
Ensembled (T=4) | 92.04 | 93.86 | 80.17 | 86.35 | 6.165 |
Small elements | |||||
Base method | 77.50 (2.35) | 91.63 (1.79) | 45.73 (2.58) | 60.60 (2.56) | |
Default TTA | 77.63 (2.28) | 90.91 (1.18) | 39.40 (2.81) | 54.38 (3.01) | |
Custom TTA | 79.10 (1.81) | 90.14 (1.43) | 51.37 (2.40) | 65.11 (1.88) | |
Ensembled (T=4) | 80.14 | 91.48 | 56.81 | 70.09 | |
