Fast Convergence for Few-shot Novel View Synthesis without Learned Priors
Additional Analysis.
Different scale of depth.
Different resolutions can capture different frequency components to achieve self-guided adaptation.
RGB
Heigh resolution depth
Middle resolution depth
Low resolution depth
RGB
Heigh resolution depth
Middle resolution depth
Low resolution depth
Visualization of novel view sampling.
In the figure below, the orange dots and lines illustrate the camera poses and view directions of the training views, while the green dots denote the average of these training poses.
The blue dots and lines represent the camera poses and view directions of the novel views we sampled.
LLFF 2 views
LLFF 3 views
LLFF 4 views
DTU 2 views
DTU 3 views
DTU 4 views
RealEstate-10K 2 views
RealEstate-10K 3 views
RealEstate-10K 4 views
Effect of the multi-scale voxel color loss.
With multi-scale voxel color loss, the model benefits from various levels of detail in the scene and leads to a better rendering result and geometry.
w/o multi-scale voxel color loss
w/ multi-scale voxel color loss
w/o multi-scale voxel color loss
w/ multi-scale voxel color loss
Effect of the self-guided adaptation.
With self-guided adaptation, the model can determine the appropriate depth through reprojection errors across different scales, significantly reducing floaters and resulting in improved geometry.
w/o self-guided adaptation
w/ self-guided adaptation
w/o self-guided adaptation
w/ self-guided adaptation
Effect of the novel view regularizations.
Novel view regularizations provide additional guidance beyond the training views, which helps prevent overfitting and ensures more accurate geometry.