The invention discloses a
data set construction and model training method and
system for the
physical form of the disabled, and the method achieves the high-precision ethical compliance
data set construction through the multi-posture modular dummy physical
simulation, high-density key point-
surface deformation joint labeling, privacy anonymization mapping and balance data enhancement strategy. The method comprises the following steps: (1) regulating and controlling a multi-freedom dummy to generate an extreme posture and limb missing sample; (2) synchronously acquiring image, video and three-dimensional size data to form a multi-
modal data packet; (3) mapping the three-dimensional data into a pixel-level dense
label; (4) optimizing marking quality and balancing sample distribution, and constructing a training
data set with strong robustness; and (5) training the incomplete figure
estimation model by using
distillation-curriculum learning. The scheme has the advantages of zero contact, wide coverage, high precision, low ethical risk and the like, and is suitable for data set generalization construction of the disabled assisting technology.