Depth map confidence estimation method based on convolutional neural network

Abstract

The invention discloses a depth map confidence estimation method based on a convolutional neural network, which is used for quality evaluation and post-processing operation of a depth map generated by a multi-view stereo matching algorithm, and comprises the following steps: calculating a truncated symbol distance function graph and a normal graph by using the depth map generated by the multi-view stereo matching algorithm; performing feature extraction on the truncated symbol distance function graph, the normal graph and the color graph by using a U-shaped network structure to obtain a feature graph; and using a convolutional long-short term memory structure, a prediction module, a refinement module and a multi-supervision method to predict the confidence of the depth map from the feature map and refine an estimation result. According to the method, quality evaluation can be carried out on the depth maps generated by various multi-view stereo matching algorithms, and the depth map confidence degree in multi-view stereo matching can be estimated robustly, so that evaluation of the multi-view stereo matching algorithms and post-processing of the depth maps are facilitated.

Description

technical field [0001] The present invention relates to the fields of computer vision and deep learning, and specifically relates to feature extraction and confidence prediction of intermediate results in multi-view stereo matching by using a convolutional neural network, so as to complete the quality assessment of depth maps. Background technique [0002] Depth image quality assessment is a basic topic in the field of computer vision. The current method based on deep learning is based on the depth image and the original color image output by monocular or binocular stereo matching, and constructs a convolutional neural network to predict a confidence map. This method The multi-modal data obtained by multi-objective stereo matching cannot be fully utilized, and low accuracy is produced due to the simple network structure. Contents of the invention [0003] The purpose of the present invention is to solve the deficiencies of the prior art, provide a depth map confidence esti...

AI Technical Summary

Problems solved by technology

[0002] Depth image quality assessment is a basic topic in the field of computer vision. The current deep learning-based method is based on the depth image and the original color image output by monocular or binocular stereo matching, and constructs a convolutional neural network to predict a confidence map. This method Cannot make full use of the multi-modal data obtained by multi-objective stereo matching, and produce low accuracy due to the simple network structure

Images