【摘要】 特征提取是基于深度学习的立体匹配中至关重要的一个部分。针对目前立体匹配网络在特征提取中造成的语义损失和匹配代价信息丢失问题,将特征金字塔网络作为立体匹配的特征提取部分,提取包含高层语义信息和多尺度信息的多通道特征;并使用改进的群组相关模块计算匹配代价,使网络包含更多的特征相似性信息,减少信息丢失,进而更加准确地重建弱纹理等病态区域。在SceneFlow、KITTI 2012和KITTI 2015双目数据集上进行测试评估,结果表明:提出的算法取得了较好精度,并且相比基准网络,在提高精度和弱纹理区域匹配效果的同时,所提算法没有增加较大计算负担。更多还原

【Abstract】 Feature extraction is a vital part of stereo matching based on deep learning. To reduce the discarded semantics and the loss of matching cost information in the feature extraction of stereo matching network, feature pyramid network was used as feature extraction module to extract multiple channel feature which contains high-level information and multi-scale information. And cost volume was built with the improved group-wise correlation module. It made the network contains more feature similarity information and reduces the loss of information. Furthermore, weak texture area was reconstructed more accurately. The evaluation is performed on the SceneFlow, KITTI 2012 and KITTI 2015 datasets, the results show that the accuracy of the proposed algorithm is significantly improved. Moreover, compared with the benchmark network, the proposed algorithm can improve the accuracy and the matching effect of weak texture region without increasing the much computational burden.更多还原