This webpage contains qualitative comparisons for material selection with baselines and ablations presented in the paper. The table below contains 13 columns displaying the input with query pixel annotated with a red square, material label along with red square marked on the selection, ground truth mask, followed by baselines, ablations, and then our model. Please zoom out to see all results in a single pane. We strongly recommend zooming out to 25% to see most results in the same pane.
Input | Material Label | Ground Truth | (Ours) DINO ViT8 backbone | (Ours) DINO ViT8 backbone, refined with KNN-Matting | DINO ViT16 backbone | DINO ViT8 backbone with reference concat | DINO ViT8 Single Block | UNet | KNN-Matting on Intrinsic Images (3 patches) | KNN-Matting on Intrinsic Images (3 patches) | KNN-Matting (3 patches) | KNN-Matting (5 patches) |
---|---|---|---|---|---|---|---|---|---|---|---|---|