Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (15): 246-252.DOI: 10.3778/j.issn.1002-8331.2012-0330

• Graphics and Image Processing • Previous Articles     Next Articles

Disparity Estimation Network with Confidence-Assisted Feature Enhancement

LIU Boqian, LIU Jiamin, CHEN Shenglun, WANG Zhihui, LI Haojie   

  1. International School of Information Science & Engineering, Dalian University of Technology, Dalian, Liaoning 116024, China
  • Online:2022-08-01 Published:2022-08-01

置信度辅助特征增强的视差估计网络

柳博谦,刘嘉敏,陈圣伦,王智慧,李豪杰   

  1. 大连理工大学 国际信息与软件学院,辽宁 大连 116024

Abstract: In the disparity estimation, interference factors such as edge blur, occlusion, low light, and low texture will produce local hard areas, resulting in a reduction in effective matching and the decrease of precision of estimation results. To solve the problem, a disparity estimation network with confidence-assisted feature enhancement is proposed. In the basic disparity estimation network, a sub-network is introduced to calculate the confidence of the intermediate disparity in the estimation process, which is used for getting low error regions to enhance the feature by multiplying the corresponding feature value. Better disparity maps can be obtained from enhanced cost volume. The network is trained on the Scene Flow dataset and compared with some methods in recent years. The experimental result shows that the proposed method works better than recent classical methods. Compared with baseline method, the end-point-error decreases by 0.032?9 pixel and the weight of pixels whose errors are larger than 1 pixel decreases by 0.4 percentage points, which proves the effectiveness of the proposed method.

Key words: stereo matching, deep learning, disparity confidence, feature enhancement

摘要: 视差估计中由于边缘模糊、遮挡、低光照以及少纹理等干扰因素会产出局部困难估计区域,导致有效匹配减少,估计结果的精度降低。为此,提出了一种置信度辅助特征增强的视差估计网络,在基础的视差估计网络中引入置信度估计子网络,为网络估计过程中间的视差结果计算置信度,利用置信度图可知误差较低的点,增强代价空间中对应位置特征值的幅度,达到辅助特征增强的目的。增强后代价空间可恢复得到更精准的视差。在Scene Flow数据集上训练网络,并和近年来的一些方法进行比较。实验结果表明,该方法优于近年来的经典方法。相比于基础网络,平均误差降低了0.032?9像素;误差大于1像素的比率降低了0.4个百分点。由此可见该方法可以有效提升视差估计网络的性能。

关键词: 立体匹配, 深度学习, 视差置信度, 特征增强