Parallel 3D Estimation via Local Hypothesis Propagation

The patent segments the 3D information calculation into independent hypothesis evaluation units for each pixel. Instead of sequentially processing pixels with dependency on preceding pixels, each pixel's hypotheses are evaluated independently using local image data and propagated hypotheses from previous iterations, enabling parallel processing across thousands of GPU cores while maintaining calculation accuracy through iterative refinement.

The patent performs preliminary action by generating and propagating hypotheses from previous frames or iterations before final selection. Hypotheses are pre-computed and propagated to current pixel positions based on motion estimates and spatial coherence, allowing the current iteration to start with informed hypotheses rather than exhaustive search, thus enabling parallelization while preserving accuracy.

The patent applies partial action by using a small set of propagated hypotheses from neighboring pixels rather than comprehensive numerical updates from all neighbors. This selective approach uses only the most relevant hypotheses (e.g., from a limited spatial neighborhood or based on motion coherence) to inform current pixel estimation, reducing computational complexity while maintaining accuracy through the sufficiency of key neighbor information.

The patent changes parameters by representing 3D information as a discrete set of hypotheses with associated confidence scores rather than continuous numerical values requiring complex optimization. This hypothesis-based representation transforms the problem from continuous parameter optimization to discrete hypothesis selection and propagation, significantly reducing computational complexity while preserving accuracy through the structured hypothesis space.

The patent segments the computation into independent pixel-level hypothesis evaluation tasks that can be executed in parallel on GPU cores. Each pixel's hypothesis evaluation uses only local image data and propagated hypotheses, minimizing the need for global memory access. This segmentation allows efficient use of GPU parallelism while reducing memory bandwidth requirements by limiting data access to locally relevant information.

The patent extracts and processes only the essential hypotheses needed for each pixel calculation, rather than accessing and processing all possible 3D information. By extracting and propagating only the most relevant hypotheses from previous iterations or neighboring pixels, the system reduces the volume of data that needs to be transferred across memory bandwidth, enabling efficient parallel processing on GPUs.