基于多任务学习的立体匹配算法

引入辅助任务信息有助于立体匹配模型理解相关知识,但也会增加模型训练的复杂度。为解决模型训练对额外标签数据的依赖问题,提出了一种利用双目图像的自相关性进行多任务学习的立体匹配算法。该算法在多层级渐进细化过程中引入了边缘和特征一致性信息,并采用循环迭代的方式更新视差图。根据双目图像中视差的局部平滑性和左右特征一致性构建了损失函数,在不依赖额外标签数据的情况下就可以引导模型学习边缘和特征一致性信息。提出了一种尺度注意的空间金字塔池化,使模型能够根据局部图像特征来确定不同区域中不同尺度特征的重要性。实验结果表明:辅助任务的引入提高了视差图精度,为视差图的可信区域提供了重要依据,在无监督学习中可用于确定单视角可见区域;在KITTI2015测试集上,所提算法的精度和运行效率均具有一定的竞争力。

Algorithm for Stereo Matching Based on Multi-Task Learning

The introduction of auxiliary task information is helpful for the stereo matching model to understand the related knowledge,but the complexity of model training increases.In order to solve the problem of dependence on extra label data during model training,we proposed an algorithm based on multi-task learning for stereo matching by using the autocorrelation of binocular images.This algorithm introduces the edge and feature consistency information in the multi-level progressive refinement process and updates the disparity map in a cyclic and iterative manner.According to the local smoothness of disparity and the consistency of left and right features of binocular images,a loss function is constructed to guide the model to learn the edge and feature consistency information without relying on additional label data.A spatial pyramid pooling with scale attention is proposed to enable the model to determine the importance of different scale features based on the local image features in different areas.The experimental results show that the introduction of auxiliary tasks not only improves the accuracy of disparity maps,but also provides a significant basis for the trusted regions of disparity maps.It can also be used to determine the single-view visible areas in unsupervised learning.The proposed algorithm has certain competitiveness in terms of accuracy and operating efficiency on the KITTI2015 test dataset.