A Real-time Stereo-Matching Algorithm for the Aggregation of Multiscale Features and Costs
Aiming at the operation speed and accuracy of stereo matching,a real-time stereo matching algorithm with multi-scale feature and cost aggregation is proposed.Firstly,a feature aggregation module can effectively extract the features of different scales.Secondly,the L1 distance is used in the matching cost calculation stage to calculate the matching cost.In the cost aggregation stage,the within-scale cost aggregation and cross-scale cost aggregation module are used to replace the calculation cost caused by regularization of the matching cost volume with 3D convolution.Finally,the disparity regression is performed,and the spatial propagation network is optimized in the disparity optimization stage.The experimental results show that the stereo matching algorithm can guarantee the accuracy and speed of matching.
NETL
NSTL
万方数据
