Table 2 Comparison of different inpainting methods in EDM completion. The FID is calculated between an ensemble of distance matrices, which are generated by the Davies-Harte algorithm, and reconstructed samples corresponding to three different sparsity values \(\mu\). The dimension of the InceptionV3 embedding used for FID is 64. The rank measures the contribution of the first \(r=5\) singular values of the reconstructed matrix in the nuclear norm. The number of matrices used in the computation of the FID metrics is equal to \(10^5\).
Sparsity | Metric | Method | ||||
|---|---|---|---|---|---|---|
RePaint | DDRM | DDNM | DDPM | Database search | ||
\(\mu =0.25\) | RMSE\(\downarrow\) | \(0.49 \pm 0.02\) | 0.170 ± 0.017 | \(0.211 \pm 0.018\) | \(0.313 \pm 0.023\) | \(1.12 \pm 0.12\) |
FID\(\downarrow\) | \(0.0446 \pm 0.0026\) | 0.013 ± 0.0017 | \(0.027 \pm 0.0015\) | \(0.0235 \pm 0.0019\) | \(1.225 \pm 0.009\) | |
Rank\(\uparrow\) | \(0.858 \pm 0.025\) | \(0.853 \pm 0.023\) | \(0.854 \pm 0.022\) | \(0.849 \pm 0.025\) | \(0.65 \pm 0.05\) | |
\(\mu =0.5\) | RMSE\(\downarrow\) | \(0.54 \pm 0.04\) | 0.241 ± 0.027 | \(0.325 \pm 0.027\) | \(0.55 \pm 0.05\) | \(1.61 \pm 0.18\) |
FID\(\downarrow\) | \(0.053 \pm 0.003\) | 0.018 ± 0.002 | \(0.053 \pm 0.002\) | \(0.0246 \pm 0.0007\) | \(1.79 \pm 0.01\) | |
Rank\(\uparrow\) | \(0.86 \pm 0.025\) | \(0.854 \pm 0.025\) | \(0.853 \pm 0.025\) | \(0.843 \pm 0.030\) | \(0.63 \pm 0.05\) | |
\(\mu =0.75\) | RMSE\(\downarrow\) | \(0.68 \pm 0.06\) | 0.42 ± 0.04 | \(0.56 \pm 0.04\) | \(1.23 \pm 0.18\) | \(1.97 \pm 0.22\) |
FID\(\downarrow\) | \(0.075 \pm 0.003\) | 0.034 ± 0.003 | \(0.116 \pm 0.002\) | \(0.096 \pm 0.003\) | \(1.25 \pm 0.011\) | |
Rank\(\uparrow\) | \(0.863 \pm 0.025\) | \(0.854 \pm 0.027\) | \(0.854 \pm 0.027\) | \(0.82 \pm 0.04\) | \(0.65 \pm 0.05\) | |
