计算机科学
人工智能
算法
计算机视觉
模式识别(心理学)
压缩传感
估计
计算几何
数学
缩小
稀疏矩阵
信号处理
噪音(视频)
估计理论
稀疏逼近
信噪比(成像)
实体造型
噪声测量
信号重构
最大似然
迭代重建
人工神经网络
作者
Yuli Wang,Weiwei Zhou,Yanping Li,Yizhang Liu
标识
DOI:10.1109/lsp.2026.3743720
摘要
Correspondence pruning identifies inliers from putative sets, playing a vital role in 3D vision. While Transformer-based methods capture global context well, they may fail to capture explicit geometric topology crucial for rigid motion consistency. Additionally, existing multi-stage networks pass features sequentially, neglecting long-range context and causing error accumulation. To address these limitations, we present CGSNet, a framework that explicitly injects spatial priors and prevents cumulative errors. Specifically, the Exphormer-infused Multi-Graph Consensus (ExMGC) block aggregates features from the pruning stages through Exphormer-style sparse graph aggregation. Concurrently, a Cross-Graph Spatial Attention (CGSA) module encodes explicit geometry by fusing spatial coordinates and motion vectors. On YFCC100M without RANSAC, CGSNet achieves 73.98% mAP5$^\circ$.
科研通智能强力驱动
Strongly Powered by AbleSci AI