3D Shape Reconstruction Using Reliability-Based Voxel Correction

The patent applies local quality by differentiating between reliable and unreliable regions in the three-dimensional shape data. It identifies portions with insufficient valid cameras and applies different processing strategies: deleting unreliable portions and using approximation models only where necessary, while preserving accurate regions. This localized approach ensures that high-fidelity regions maintain their precision while low-fidelity regions are appropriately handled, resolving the contradiction between completeness and accuracy.

Instead of traditionally adding missing parts through complementation (which causes inflation), the patent inverts the approach by selectively deleting unreliable portions and using approximation models only as a fallback. This inversion prevents excessive inflation by removing the problematic complementation step and replacing it with a deletion-and-approximation strategy that maintains shape fidelity.

The patent applies local quality by evaluating each portion of the three-dimensional shape data based on the number of valid cameras that captured it. Regions with sufficient viewpoints are processed with high accuracy using Silhouette Volume Intersection, while regions with insufficient viewpoints are identified and handled differently through deletion or approximation models. This localized differentiation maintains processing speed while improving overall shape accuracy.

The patent applies partial action by selectively applying approximation models only to portions with insufficient valid cameras, rather than applying complementation globally. This partial application of approximation prevents excessive inflation in well-captured regions while providing necessary filling only where viewpoints are insufficient, thus maintaining both speed and accuracy.

The patent addresses uneven camera distribution by applying local quality assessment to each portion of the object. It identifies regions where valid cameras are insufficient and handles them differently from well-covered regions. Approximation models are applied selectively to under-covered portions, ensuring that coverage is maintained while fidelity is preserved in well-captured areas.

The patent uses parameter changes by introducing a threshold for the number of valid cameras required for reliable reconstruction. Portions exceeding this threshold are processed with high fidelity, while those below the threshold trigger approximation model application. This parameter-based differentiation resolves the contradiction between coverage and fidelity by adapting processing quality to local camera distribution.