Comparison

Col. 3 shows the results of Gilet et al. [2010]. These results were generated by the authors of [Gilet et al. 2010] themselves, who reported that significant manual intervention and trial and error were needed to achieve these results, while our approach is automatic. Additionally, their results typically exhibit a lack of high-frequency detail (see, e.g., 4, 10, 12), increased regularity (see, e.g., 6, 16, 19), and/or subtle color shifts (see, e.g., 2, 6, 16). Note that, compared to our approach, Gilet et al. additionally use cosines and histogram matching, which can improve the results in some cases (see, e.g., 3, 9). Although our approach is compatible with these additions, we have chosen not use them, since the usage of cosines and histogram matching is not compatible with surface Gabor noise and analytic filtering of Gabor noise.

Col. 4 shows the results of Lagae et al. [2010]. Their method cannot handle anisotropy (see, e.g., 2, 3 and 4).

Col. 5 shows the results of Heeger and Bergen [1995]. Their method cannot handle anisotropy well (see, e.g., 7, 13, and 16).

Col. 6 shows the results of Lefebvre and Hoppe [2005]. Their method is not procedural, and cannot preserve the anisotropic detail well in all cases (see, e.g., 4, 15, 19).

(1) Exemplar
(2) Ours
(3) Gilet et al.
(4) Lagae et al.
(5) Heeger et al.
(6) Lefebvre et al.
1: ConcreteBare0280_39_thumbhuge_128
2: ConcreteFloors0068_2_thumbhuge_128
3: fabric_2002
4: fabric_2004
5: FabricPlain0001_thumbhuge_128
6: FabricPlain0020_2_thumbhuge_128
7: FabricPlain0059_5_thumbhuge_128
8: Farmland0008_19_thumbhuge_128
9: FloorsRounded0041_2_thumbhuge_128
10: Leather0002_2_thumbhuge_128
11: MarbleGranite0007_5_thumbhuge_128
12: MetalBare0131_thumbhuge_128
13: paint_2001
14: paint_2005
15: PaperDecorative0028_6_thumbhuge_128
16: PaperDecorative0048_2_thumbhuge_128
17: RockSediment0006_7_thumbhuge_128
18: wall_2005
19: WoodBurned0068_1_thumbhuge_128