We ablate the key components in our framework by comparing against an image-space L2 loss (col 2) and showing our result before (col 3) and after (col 4) refinement.
Due to space-constraints, we don't show an image-space L1 loss (which performs similarly to an image-space L2 loss), or volumetric L1 and L2 losses (which also underperform our correlation-based volumetric loss).
However, in the accompanying PDF we provide a quantitative comparison also to these additional image-space and volumetric losses.
| Input | Image space L2 | Ours unrefined | Ours refined | |
|---|---|---|---|---|
A kangaroo wearing a christmas sweater |
|
|
|
|
| | ||||
A dog wearing big sunglasses |
|
|
|
|
| | ||||
A cat wearing a birthday party hat |
|
|
|
|