3D Human Dynamic Texture Mapping With Neural Image Reconstruction

The patent segments the texture map generation process into multiple training iterations, where a neural network is trained progressively on subsets of image data. Instead of processing all high-resolution images at once, the system divides the training into batches, reducing the computational load per iteration while accumulating quality improvements over multiple passes. This allows high-quality texture generation without overwhelming processor power requirements.

The patent applies partial action by using a portion of available image data for training rather than requiring complete high-quality datasets. The neural network is trained on selected subsets of images, and the texture map quality improves iteratively with each training batch. This approach achieves high-quality results without needing to process excessive amounts of data simultaneously, reducing both computational power and memory bandwidth requirements.

The patent performs preliminary action by pre-training the neural network on available image data before final texture map generation. The system prepares the neural network model in advance through multiple training iterations on training images, so that when actual texture mapping is needed, the pre-trained model can quickly generate high-quality results. This preliminary training phase separates the computationally intensive work from the final generation time, reducing overall production time.

The patent uses copying by generating texture maps through neural network synthesis rather than direct photographical mapping. The neural network learns to copy and reconstruct texture patterns from training images, creating photorealistic texture maps that replicate the appearance of high-quality photographs without requiring actual high-resolution source images for every view. This copying approach through learning reduces both time and cost compared to conventional photogrammetry methods.