Depth Map Upsampling Using Adaptive Support Windows

The patent segments the support window into multiple sub-windows based on depth discontinuity detection. By dividing the filtering region into smaller segments where depth is relatively uniform, the computational complexity is reduced while maintaining precision. Each sub-window can be processed independently with simpler operations, avoiding the need to compute bilateral weights across the entire large window.

The patent performs preliminary detection of depth discontinuities and identifies support regions before executing the up-sampling operation. By pre-segmenting the image based on depth edges and preparing sub-window structures in advance, the actual filtering process becomes computationally lighter. This preliminary analysis allows subsequent filtering to operate on smaller, pre-defined regions rather than computing weights dynamically across large windows.

The patent divides large windows into smaller sub-windows based on depth discontinuity boundaries. This segmentation allows the system to effectively utilize large window sizes for capturing sufficient contextual information while reducing the actual computational burden by processing smaller sub-regions independently. The segmentation strategy enables large effective support regions without the quadratic computational cost of processing the entire large window.

The patent applies different filtering strategies to different local regions based on depth characteristics. In regions with depth discontinuities, smaller sub-windows are used to preserve edges, while in smooth regions, larger effective coverage is achieved through adaptive combination of sub-window results. This local adaptation optimizes the balance between window size benefits and computational cost for each specific region.

The patent applies different filtering behaviors to different local regions based on actual depth discontinuity detection. In regions where depth discontinuities are detected at intensity edges, the filtering preserves edges similar to bilateral filters. In regions where intensity edges do not correspond to depth discontinuities, the filtering allows smooth variations by adapting the support region boundaries. This local differentiation resolves the contradiction by making the edge preservation behavior context-dependent rather than universal.

The patent makes the support region boundaries dynamic rather than fixed, allowing them to adapt to local depth characteristics. The support regions are adjusted based on detected depth discontinuities, expanding in smooth regions and contracting at actual depth edges. This dynamic adaptation enables the filter to preserve edges where needed while allowing smooth variations where intensity edges do not correspond to depth boundaries, improving overall up-scaling reliability.