The invention belongs to the technical field of
computer vision and large-scale three-dimensional geometric
data labeling, and relates to a three-dimensional model outer surface automatic labeling method based on
image segmentation mapping. The method comprises the following steps: S1, importing a triangular
mesh model to be labeled, arranging N virtual cameras, and outputting N two-dimensional color images and depth maps; s2, inputting the two-dimensional
color image into a pre-trained high-performance semantic segmentation network, and outputting a semantic tag of a two-bit pixel point and a quantized segmentation confidence coefficient; s3, projecting the center point of the three-dimensional network patch to a two-dimensional pixel coordinate by using camera parameters, and performing depth consistency
verification; s4, collecting all effective labels and confidence coefficients for the patches observed by the multiple visual angles, and determining a final
label and a comprehensive confidence coefficient; and S5, automatically locking the high-reliability area, and marking the low-reliability or conflict area as a special
label to be rechecked. According to the method, the efficiency, precision and consistency of large-scale three-dimensional model
semantic annotation are remarkably improved.