Single-view multi-person body reconstruction method based on depth UV prior

Abstract

The invention provides a single-view multi-person body reconstruction method based on depth UV prior, and the method positions and segments a body image through employing a two-dimensional posture estimation result, and carries out the dynamic reconstruction of a plurality of body three-dimensional grid models from the segmented image in combination with the UV body prior based on a variational auto-encoder. The method comprises the following steps: constructing a human body UV prior network based on a variational auto-encoder; training a human body reconstruction network; and using the human body reconstruction network to complete multi-person human body three-dimensional reconstruction. According to the method, the purity of the image information input into the network is ensured, the generalization ability of the network is enhanced, extra human body information can be provided when the network reconstructs a complete human body three-dimensional model from a small amount of visible human body image information, and the rationality of the three-dimensional human body model obtained through reconstruction is ensured, The method can support the reconstruction of a part of the human body region in the image under the condition that the human body region is shielded and invisible, and can realize the reconstruction of a three-dimensional human body model with an absolute position under the condition that multiple persons are shielded.

Description

technical field [0001] The invention belongs to the technical field of computer three-dimensional vision, and relates to a multi-person human body reconstruction method based on depth UV prior. Background technique [0002] Multi-person reconstruction technology based on a single RGB image plays an important role in virtual reality applications such as sports auxiliary training, group behavior analysis, and holographic communication. For a long time, the reconstruction of high-precision human mesh models has relied on complex and expensive multi-view stereo systems. In recent years, with the rapid development of deep learning technology, the data-driven single-view human body reconstruction method has become the mainstream solution in the field of human body reconstruction due to its accuracy and operating efficiency. However, in multi-person scenes, the occlusion and interaction between people lead to ambiguity at the pixel level of the image. It is difficult for single-pe...

AI Technical Summary

Problems solved by technology

This method avoids the interference in the image, but it is difficult to accurately reconstruct a complete 3D human body with a small amount of image information for severely occluded objects

Images