Implicit Scene Reconstruction With SVD Plane Regularization

The patent modifies the NeRF objective function by introducing a plane regularization term that constrains the geometry parameters. This regularization term enforces planarity constraints on flat surfaces during training, preventing overfitting while maintaining view synthesis capability. The parameter change is applied to the loss function to balance photometric accuracy with geometric correctness.

The patent implements a feedback mechanism through the plane regularization term that continuously monitors and corrects geometric deviations during training. The regularization term provides feedback to the neural network about how much the reconstructed surfaces deviate from expected planar geometry, allowing the model to self-correct and improve geometry reconstruction accuracy while preserving view synthesis performance.

The patent inverts the approach of RegNeRF by not forcing surface normals to be orthogonal to central rays. Instead, it applies plane regularization that allows surface normals to vary while maintaining overall planarity of flat surfaces. This inversion of the constraint approach preserves both depth smoothness and accurate geometric representation of flat surfaces without the orthogonal normal constraint.

The patent applies different regularization strategies to different regions of the scene. Flat surfaces receive plane regularization constraints that enforce local planarity, while other regions maintain the original NeRF flexibility. This local quality approach allows accurate geometry reconstruction where needed without over-constraining the entire scene, resolving the contradiction between smoothness and geometric accuracy.