Table 7 Performance with varying kernel sizes on the dataset of Tanzania.
From: Fine-grained population mapping from coarse census counts and open geodata
Training setting | Kernel size | #params (k) | \(R^2 \uparrow\) [%] | MAE \(\downarrow\) | MAPE \(\downarrow\) [%] |
|---|---|---|---|---|---|
Coarse | \(1 \times 1\) (ours) | 34 | 85.7 ± 0.9 | 3100 ± 40 | 21.6 ± 0.2 |
\(3 \times 3\) | 313 | 83.5 ± 1.3 | 3700 ± 110 | 26.6 ± 0.7 | |
\(5 \times 5\) | 870 | 80.9 ± 1.0 | 3900 ± 100 | 28.4 ± 0.8 | |
Fine | \(1 \times 1\) (ours) | 34 | 87.6 ± 0.2 | 2890 ± 20 | 20.4 ± 0.3 |
\(3 \times 3\) | 313 | 86.0 ± 2.1 | 3010 ± 100 | 20.9 ± 0.5 | |
\(5 \times 5\) | 870 | 86.8 ± 1.0 | 3030 ± 50 | 21.5 ± 0.4 |
