.png){kind=icon}
.png){kind=icon}
Return
PRNet: Parallel Reinforcement Network for two-view correspondence learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.112978.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Two-view correspondence learning is a fundamental task in computer vision for locating the same object across two different images, and its essence lies in capturing the context from the images. Recent studies employ the standard Convolutional Neural Network (CNN) as the core architecture to capture the context. However, the inherent nature of the CNN's local receptive field and pooling operations can result in the loss of certain semantic context. This can result in the CNN-based correspondence learning methods having an insufficient understanding of the global context, especially on image pairs including challenges like significant viewpoint changes, repetitive structures and weak textures. To address this issue and these challenges, we propose a novel correspondence learning method called Parallel Reinforcement Network (PRNet). Firstly, we design a reinforcement injection block, not only to dynamically refine feature weights by using channel attention mechanism, but also to preserve important details for alleviating over-smoothing issue by strengthening the network's capacity. Secondly, to alleviate the potential issue of overlooking the weak local context by CNN, we propose a parallel fusion block to integrate both shallow and deep features, preserving local details and enhancing global context. We evaluate the performance of the proposed PRNet on an outlier rejection task and a relative pose estimation task. The experimental results demonstrate the proposed PRNet exceeds several existing state-of-the-art methods in various challenging scenarios.
Keywords:
Two-view correspondence
Channel attention mechanism
Parallel fusion
Pose estimation
Outlier rejection
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W