Keskeiset käsitteet
A neural signed distance field (SDF) 3D reconstruction framework with view-dependent normal compensation that enhances the quality of indoor scene reconstructions by addressing multi-view inconsistency between monocular normal priors.
Tiivistelmä
The paper presents NC-SDF, a neural SDF 3D reconstruction framework that focuses on enhancing indoor scene reconstruction by addressing multi-view inconsistency between monocular normal priors. The key contributions are:
View-dependent normal compensation model: The framework integrates view-dependent biases in monocular normal priors into the neural implicit representation of the scene, enabling adaptive compensation for the biases and resulting in globally consistent and locally detailed reconstructions.
Informative pixel sampling strategy: An informative pixel sampling strategy is proposed to pay more attention to intricate geometry by prioritizing pixels with higher information content, further enhancing the reconstruction of geometric details.
Hybrid geometry modeling: A hybrid geometry model is introduced, combining the strengths of multi-layer perceptrons (MLPs) and voxel grids to strike a balance between modeling low-frequency structures and high-frequency details.
Comprehensive experiments on both synthetic and real-world datasets demonstrate that NC-SDF achieves state-of-the-art performance in indoor scene reconstruction, outperforming existing methods in terms of reconstruction quality.
Tilastot
The paper reports the following key metrics:
Accuracy (Acc): Lower is better
Completeness (Comp): Lower is better
Precision (Prec): Higher is better
Recall (Recall): Higher is better
F-score (F-score): Higher is better
Lainaukset
"Our NC-SDF excels in capturing intricate geometry while producing smooth surfaces in texture-less regions."
"The combination of our three designs ensures consistent and smooth surfaces while enabling sharp details in the reconstructions."