Bare Body Shape Estimation From Concealed 3D Scans

The patent creates a virtual copy of the dressed body scan by generating a synthetic dressed scan from a bare body shape through simulated clothing deformation. This virtual copy is then used to train a neural network, replacing the need for expensive optimization-based fitting during actual body shape estimation. The neural network learns the inverse mapping from dressed to bare body shapes through this synthetic training data approach.

The patent extracts and isolates the clothing deformation component by separately modeling the bare body shape and the clothing layer. Through this separation, the method extracts only the necessary information (the deformation field caused by clothing) to train the neural network, simplifying the overall data processing pipeline while maintaining accuracy.

The patent introduces a neural network model as an intermediary that processes dressed body scans and infers the underlying bare body shape without direct exposure. This intermediary learns the relationship between dressed and bare body shapes through synthetic training data, enabling accurate measurement while maintaining user privacy and convenience by allowing scans to be performed with clothing on.

Instead of requiring difficult physical replication of motion sequences with and without clothes, the patent creates a virtual copy of the motion sequence by applying the same motion transformations to both the bare body shape and the dressed body shape in simulation. This synthetic copying approach generates perfectly aligned training data without the practical difficulties of physical data collection.