一种基于证据深度学习的双目匹配不确定性估计方法

• Format: Patent
• Language: Chinese
• Published: 15.04.2025
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

本发明公开一种基于证据深度学习的双目匹配不确定性估计方法，通过卷积神经网络提取和聚合特征，进而获得一个匹配代价体和三个不确定性体，在匹配代价的指导下，计算出证据分布的四个超参数γ，ν，α和β；最后通过四个超参数计算双目匹配视差值，偶然不确定性和认知不确定性。本发明很好地反映出双目匹配的难易程度，提升不确定性估计的表现，而且面对分布外的数据时能给出较高的认知不确定性。 The invention discloses a binocular matching uncertainty estimation method based on evidence deep learning, and the method comprises the steps: obtaining a matching cost body and three uncertainty bodies through the extraction and aggregation of features through a convolutional neural network, and calculating four hyper-parameters gamma, v, alpha and beta of evidence distribution under the guidance of the matching cost; and finally, calculating a binocular matching parallax value, accidental uncertainty and cognitive uncertainty through the four hyper-parameters. According to the method, the difficulty degree of binocular matching is well reflected, the performance of uncertainty estimation is improved, and relatively high cognitive uncertainty can be given for data outside distribution.