Point-GAT:图注意力机制与关键点检测的融合

Point-GAT: The Fusion of Graph Attention Mechanism and Key Point Detection

针对以往基于关键点的目标检测存在小尺度上的检测结果不太理想,忽略关键点之间的类别语义信息的问题,提出了一种新的关键点检测算法Point-GAT。该算法通过在Hourglass和ResNeXt主干网络上加入快捷连接,解决网络深度增加带来的学习退化问题;使用反卷积和特征融合增强小尺度目标的检测效果;同时算法使用了图注意力机制,通过构建有向有权重图映射类别之间的语义关系,获得关键点之间的类别语义信息;在优化定位和回归函数的同时,加入分类损失函数分支来反映类别语义信息。在COCO数据集上实验结果表明,该算法平均精度达到了48.3%,在PASVAL VOC 2007和PASVAL VOC 2012数据集上平均精度均高于其他算法。

This paper proposes a new key point detection algorithm,Point-GAT,addressing the challenge of suboptimal key point detection results at a small scale and the oversight of category semantic information between key points in previous methods.This algorithm joins fast connections to solve the learning degradation problem caused by a too deep network.It enhances the detection effect of small-scale objects by using deconvolution and feature fusion on Hourglass and ResNeXt backbone networks.Simultaneously,the algorithm employs the graph classification method to depict the semantic relationship between categories.This is achieved by constructing a directed weighted graph to capture the category semantic information among key points.During the optimizatipon of the positioning and regression function,the branch of the classification loss function is incorporated to reflect the category semantic information.The experimental results indicate that the average accuracy of this algorithm on the COCO dataset reaches 48.3%,and the average accuracy on the PASVAL VOC 2007 and PASVAL VOC 2012 datasets is higher than that of other algorithms.