Dense correspondence prediction method based on non-rigid point cloud

Abstract

The invention discloses a dense correspondence prediction method based on non-rigid point clouds. The method comprises the following steps: respectively extracting geometric features of a three-dimensional template and point clouds by using a graph convolutional neural network and a plurality of set abstraction layers; utilizing a global regression network to deduce global displacement according to the associated global features of the template and the point cloud; utilizing a local feature embedding technology, introducing an attention mechanism, and fusing local depth features of the point cloud and geometric features of the graph; utilizing a local regression network to predict displacement increment; utilizing a weak supervision fine tuning method to process the real point cloud, and unifying the real point cloud and the two-stage regression network in a complete framework. The local geometric features of the point cloud are fully utilized, the attention strategy is adopted to improve the corresponding precision, the weak supervision fine tuning method is adopted to robustly process the real point cloud, and the conditions that the prediction model is unreasonably distorted and is obviously inconsistent with the input shape due to the lack of training data are effectively improved.

Description

technical field [0001] The invention belongs to the field of three-dimensional reconstruction, in particular to a dense corresponding prediction method based on non-rigid point cloud. Background technique [0002] Estimating dense correspondences of 3D shapes is one of the fundamental problems of computer vision and computer graphics, and an essential part of many promising applications such as gaming, robotics, and virtual reality. As 3D point clouds become more and more common, dense correspondence estimation of non-rigid point clouds plays an important role in many research topics such as multi-view stereo, object retrieval, 3D reconstruction, motion tracking, etc. However, estimating dense correspondences from point clouds remains challenging due to factors such as changes in deformed objects, incomplete 3D data, and changes in camera viewpoints. Most approaches register template models to input point clouds via non-rigid deformation techniques to obtain dense point cor...

AI Technical Summary

Problems solved by technology

However, as stated in 3DCODED, the quality of the initial models (their network predictions) is critical for deformation optimization, and unreliable initial models can cause deformation optimization to fail.

Images