Given a single RGB image, recover the pixel-wise depth.
LSIV_RMSE: It measures the surface-wise Euclidean distance between the predicted and ground-truth 3D point cloud. It is invariant to surface-wise depth scaling and translation.
WKDR: The percentage of point pairs that have incorrect depth order in the predicted depth.
For a more detailed explanation of the above metrics, please refer to the OASIS paper.
Method | LSIV_RMSE (Lower is better) | WKDR (Lower is better) |
---|---|---|
ResNetD (ImageNet + OASIS)[1] | 0.37 | 32.62% |