A Deep Learning Semi-supervised Method for Semi-dense Disparity Map Repair
The disparity map generated by the current mainstream stereo matching methods,such as BM and SG-BM,is semi-dense,which cannot be used in some scenarios requiring dense disparity map,so it must be repaired.However,the precision of traditional repair method is limited and cannot meet the requirements of high-precision scene.To solve this problem,a semi-supervised repair method based on deep learning is proposed in this paper.Based on the traditional stereo matching method,this method makes use of the advantages of deep learning in fea-ture extraction to repair the missing areas of the disparity map.The experimental results show the following points:1)The accuracy of the semi-supervised repair method is much higher than that of the traditional repair method.In the experiment,EPE and 3PE are reduced by 33.98%and 17.83%respectively compared with those of the tradi-tional repair method.2)The training results of virtual scene can be transferred to real scene to further improve the repair accuracy,EPE and 3PE can be reduced by 5.17%and 12.58%respectively,and convergence can be accel-erated,which has important practical value.
万方数据
维普
