Table 2.
Single-category pixel accuracy and intersection over union mean of different functional units configured (CBAM identifies channel attention and spatial attention modules, and RCS indicates rare class sampling strategy).
| Categories | BaseModel | +CBAM | +CBAM+RCS | |||
|---|---|---|---|---|---|---|
| acc | IoU | acc | IoU | acc | IoU | |
| road | 92.37 | 86.40 | 93.42 | 87.33 | 92.35 | 84.91 |
| glass | 90.96 | 84.25 | 93.36 | 86.14 | 92.44 | 85.58 |
| tree | 96.12 | 91.38 | 97.2 | 92.0 | 96.56 | 90.64 |
| high vegetation | 79.55 | 67.91 | 84.09 | 73.82 | 85.63 | 83.18 |
| object | 65.18 | 58.19 | 79.62 | 67.15 | 81.56 | 68.17 |
| building | 92.87 | 82.93 | 93.29 | 85.51 | 93.17 | 85.12 |
| log | 57.13 | 42.71 | 60.15 | 47.92 | 63.28 | 53.74 |
| sky | 87.66 | 79.83 | 89.79 | 79.86 | 85.42 | 78.14 |
| rock | 68.08 | 63.85 | 70.08 | 65.12 | 66.22 | 68.62 |
| water | 83.52 | 74.29 | 85.03 | 77.05 | 85.47 | 79.41 |