3D Model Segmentation Method Based on Fusion of Multi-view Features by Projective Fully Convolutional Network

Abstract

The invention discloses a projection full convolution network three-dimensional model segmentation method based on fusion of multi-view features, comprising: step 1, collecting data from an input three-dimensional grid model data set; step 2, using an FCN full convolution network that fuses multi-view features Carry out semantic segmentation on the model projection rendering image, and obtain the probability that the pixels of the model projection rendering image in each viewpoint direction are predicted to be each label; step 3, reverse-project the semantic segmentation probability map of the model projection rendering image in each viewpoint direction and merge The maximum view pooling is used to obtain the probability that the model patch is predicted as each label; step 4, the Graph Cut algorithm is used for optimization to obtain the final predicted label of the model patch.

Description

technical field [0001] The invention belongs to the fields of computer image processing and computer graphics, and in particular relates to a projection full convolution network three-dimensional model segmentation method based on fusion of multi-view features. Background technique [0002] In recent years, with the emergence of more and more 3D modeling software, and depth sensors, such as Kinect, etc., have been widely used on platforms for collecting depth data, 3D model data has exploded on the Internet, 3D There are also a large number of representations of the model, such as point clouds, voxels, and patches. This trend makes the analysis of 3D models a hot research field. At present, the research in the field of image analysis has achieved fruitful results, and the introduction of deep learning framework has further improved the effect. However, convolution operations on 2D images cannot be directly applied to 3D models, making it difficult to apply deep learning me...

AI Technical Summary

Problems solved by technology

[0007] 4. Analysis methods must be robust to noise, downsampling, and the diversity of similar models

However, the choice of viewing angle of this method is too time-consuming, and the optimization of CRF training at the end prolongs the training time of the entire framework, which is less practical.

Images