.png){kind=icon}
返回
Feature enhancement network for stereo matching
.png){kind=icon}
PRE
AI
DOI：10.1016/j.imavis.2022.104614.png)
.png){kind=icon}
.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
In general stereo matching methods, there are mainly two kinds of weaknesses: the interior of large objects with unclear texture; object boundaries and small objects. One of the reasons for the first case is the underuse of inter -feature and intra-feature contexts, leading to the lack of discriminability of feature representations. And the sec-ond case could be caused result from the loss of fine-grained structural information in the features. In this work, we present a novel stereo matching network to enhance the feature representations, which contains two prom-inent components: pyramid context enhancement module and cross-modal enhancement module. Pyramid con-text enhancement module first constructs multi-scale pyramid features to promote global representation, and then exploits the inter-feature and intra-feature contextual information to improve the discriminability of fea-tures. Cross-modal enhancement module proactively degrades features of RGB image and disparity map to im-plicitly extract structures, and fuses both structures to recover the modal consistent structural information of features. Moreover, we design a disparity weight loss function which can calculates the adaptive weight accord-ing to prediction error to balance the loss response. Extensive experiments demonstrate that our method boosts the performance, and achieves better results on five challenging datasets than other typical methods. Code is available at: https://github.com/csl1994/FENet. (c) 2023 Elsevier B.V. All rights reserved.
Keyword:
Stereo matching
Deep learning
Pyramid feature representation
Cross -modal fusion
Disparity weight loss
期刊
IF:
4.2
论文数:
4.1K
被引数:
6.7K
机构
引用论文
.png){kind=icon}
分享收藏分享收藏分享收藏
查看更多内容