Unsupervised dense matching method and system based on reconstruction mapping consistency

• Main Authors: LIN YUZHUN, RUI JIE, WEI LINSU, WANG JIANFENG, LI HUA, MIAO YUZHE, GUO HAOJUN, WANG FAN, GUAN KAI, WANG SHUXIANG, JIN FEI, LIU ZHI, GAO XUEMEI, LIU XIAO
• Format: Patent
• Language: Chinese; English
• Published: 10.12.2021
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; HANDLING RECORD CARRIERS; PHYSICS; PRESENTATION OF DATA; RECOGNITION OF DATA; RECORD CARRIERS

The invention belongs to the technical field of dense matching, and relates to an unsupervised dense matching method and system based on reconstruction mapping consistency. The method comprises: constructing a dense matching network, and learning mapping from an input image to a disparity map by using an unsupervised loss function as a target constraint function, wherein the unsupervised loss function comprises a reconstruction mapping consistency loss function, a smooth loss function and a left-right consistency loss function; collecting scene sample data, and dividing the scene sample data into a training sample and a test sample; pre-training the dense matching network by using the training sample, and performing test optimization on the pre-trained network by using the test sample; and performing dense matching of the target scene data by using the dense matching network after test optimization. Reconstruction mapping consistent loss is utilized, smooth and left-right consistent loss is combined to serve