Table 4 Comparison of different inference phase strategies on the PLM validation set
From: Automatic damage identification of Sanskrit palm leaf manuscripts with SegFormer
Method | Hit (%) | class IoU (%) | mHit (%) | mIoU (%) | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
INC | BRE | FIB | CON | IMP | INC | BRE | FIB | CON | IMP | |||
None | 76.8 | 71.2 | 71.6 | 54.1 | 97.8 | 57.0 | 42.7 | 43.4 | 33.8 | 93.5 | 54.1 | 71.6 |
TTA | 77.9 | 71.2 | 71.7 | 54.5 | 97.8 | 57.4 | 43.3 | 43.4 | 35.3 | 93.2 | 54.5 | 71.7 |
TLC | 76.8 | 72.3 | 70.8 | 54.1 | 99.6 | 57.5 | 42.8 | 42.8 | 33.5 | 93.9 | 54.1 | 70.8 |
TTA + TLC | 79.6 | 74.7 | 72.9 | 54.9 | 98.4 | 59.3 | 43.3 | 42.9 | 35.1 | 93.9 | 54.9 | 72.9 |
TTA + TLC + Post | 79.6 | 74.8 | 73.0 | 54.8 | 98.4 | 59.8 | 42.3 | 42.5 | 35.2 | 94.1 | 54.8 | 73.0 |